Program of ML in PL Conference 2025

/ Contributed talks

Adrian Łańcucki

NVIDIA

Contributed talk 1: Learning Dynamic Segmentation and Compression of Sequences in Transformer LLMs

Thursday / 16 October 9:30 - 10:00 Main Hall (CfC Session 1)

slides slides (online preview)

Abstract:

Transformer-based LLMs excel at language tasks, but their efficiency hinges on input sequence length. Typically, input resolution—imposed by a tokenizer—remains unchanged across all layers. In this talk, we introduce methods that enable end-to-end learning to dynamically pool, compress, or sparsify input or key-value token sequences. Our adaptive pooling methods enable training character-level models that internally construct and operate on word-like segments. Additionally, these methods allows to track down and remove redundancies, resulting in substantial performance gains during training or inference. Finally, we arrive at a surprisingly practical method—Dynamic Memory Sparsification—that not only significantly improves latency and throughput by compressing the KV cache but also boosts accuracy, as demonstrated across several reasoning tasks.

Biography:

Adrian Łańcucki is a senior engineer at NVIDIA, where he optimizes the performance of LLMs and conducts research on representation learning, unsupervised segmentation, and generative modeling for text and speech. He is the author of FastPitch, a widely adopted text-to-speech model that significantly accelerated speech synthesis research. In 2019, Adrian obtained a Ph.D. in machine learning from the University of Wroclaw, Poland, and he has since actively collaborated with academia.

Łukasz Borchmann

Snowflake

Co-authors:

Michał Pietruszka, Wojciech Jaśkowski, Dawid Jurkiewicz, Piotr Halama, Paweł Józiak, Łukasz Garncarek, Paweł Liskowski, et al.

Contributed talk 2: State-of-the-Art Document AI on a Single 24GB GPU

Thursday / 16 October 10:00 - 10:30 Main Hall (CfC Session 1)

slides

Abstract:

The vast portion of workloads employing LLMs involves answering questions grounded on PDF or scanned content. We introduce the Arctic-TILT achieving accuracy on par with models 1000× its size on these use cases. It can be finetuned and deployed on a single 24GB GPU, lowering operational costs while processing rich documents with up to 400k tokens. The model establishes state-of-the-art results on seven diverse Document Understanding benchmarks, as well as provides reliable confidence scores and quick inference, essential for processing files in large-scale or time-sensitive enterprise environments. We release Arctic-TILT weights and an efficient vLLM-based implementation on a permissive license.

Biography:

Lukasz is a researcher specializing in natural language processing and document understanding. With a strong background in the industry and several international competition wins, he has contributed to the advancement of language modeling, particularly in multimodal models incorporating visual and layout features alongside textual information. He came to Snowflake as part of the Applica.ai acquisition and was recently involved in developing Snowflake Arctic and Arctic-TILT LLMs. His PhD thesis focused on neural network architectures shifting paradigms toward what is now called generative AI.

Jakub Krajewski

University of Warsaw

Co-authors:

Marcin Chochowski, Daniel Korzekwa

Contributed talk 3: Scaling Fine-Grained MoE Beyond 50B Parameters: Empirical Evaluation and Practical Insights

Thursday / 16 October 10:30 - 11:00 Main Hall (CfC Session 1)

slides slides (online preview)

Abstract:

Mixture of Experts (MoE) architectures have emerged as pivotal for scaling Large Language Models (LLMs) efficiently. In a previous paper authored by me and other researchers at IDEAS NCBR, we proposed granularity hyperparameter and derived Scaling Laws for Fine-Grained MoE. Since then, fine-grained MoE approaches - utilizing more numerous, smaller experts - have demonstrated potential in improving model convergence and quality. This work, done during an internship at NVIDIA Warsaw with Marcin Chochowski and Daniel Korzekwa, proposes a set of training recipes and provides a comprehensive empirical evaluation of fine-grained MoE, directly comparing its scaling properties against standard MoE configurations for models with up to 56B total (17B active) parameters. We investigate convergence speed, model performance on downstream benchmarks, and practical training considerations across various setups. Overall, at the largest scale we show that fine-grained MoE achieves better validation loss and higher accuracy across a set of downstream benchmarks. This study offers empirical grounding and practical insights for leveraging fine-grained MoE in the development of future large-scale models.

Biography:

Jakub Krajewski is a PhD student working on LLM efficiency, scaling laws and Mixture of Experts.

Adam Pardyl

Jagiellonian University; IDEAS NCBR

Co-authors:

Dominik Matuszek, Mateusz Przebieracz, Marek Cygan, Bartosz Zieliński, Maciej Wołczyk

Contributed talk 4: FlySearch: Exploring how vision-language models explore

Thursday / 16 October 9:30 - 10:00 Hall A (CfC Session 2)

slides

Abstract:

The real world is messy and unstructured. Uncovering critical information often requires active, goal-driven exploration. It remains to be seen whether Vision-Language Models (VLMs), which recently emerged as a popular zero-shot tool in many difficult tasks, can operate effectively in such conditions. In this paper, we answer this question by introducing FlySearch, a 3D, outdoor, photorealistic environment for searching and navigating to objects in complex scenes. We define three sets of scenarios with varying difficulty and observe that state-of-the-art VLMs cannot reliably solve even the simplest exploration tasks, with the gap to human performance increasing as the tasks get harder. We identify a set of central causes, ranging from vision hallucination, through context misunderstanding, to task planning failures, and we show that some of them can be addressed by finetuning. We publicly release the benchmark, scenarios, and the underlying codebase.

Biography:

Adam Pardyl is a PhD candidate in Group of Machine Learning Research at Jagiellonian University and a researcher at IDEAS NCBR. His research interests include computer vision for embodied AI, reinforcement learning, and MLLMs for robotics.

Gracjan Góral

University of Warsaw

Co-authors:

Emilia Wiśnios, Paweł Budzianowski

Contributed talk 5: How Good Are Open-Source Models for Robot Learning?

Thursday / 16 October 10:00 - 10:30 Hall A (CfC Session 2)

Abstract:

The reliance on large, curated datasets is a primary obstacle to progress in robotics. Generative Value Learning (GVL) offers a promising direction by using vision-language models (VLMs) to estimate task progress for self-supervised learning. However, the use of closed-source models limits widespread research and deployment. In this work, we investigate the viability of open-source VLMs as a foundation for GVL. Our empirical evaluation reveals that while a discernible performance gap to leading proprietary models exists, current open-source alternatives provide a promising and accessible foundation for this task. To build on this foundation, we demonstrate how to fine-tune these VLMs natively on the GVL task, which improves their capacity for physical reasoning. Finally, to standardize evaluation, we propose a new benchmark suite covering a diverse set of manipulation scenarios.

Biography:

Former math student, now exploring the boundaries between mathematics, artificial intelligence, and language. My work focuses on language models and their ability to reason, reflect, and sometimes hallucinate. I am fascinated by the intersection of psychology and AI. In particular, I apply psychological frameworks and experimental paradigms to study, challenge, and sometimes surprise artificial models. My research is rarely done alone – I share my home office with five cats ("the demons"), who are convinced every keyboard was made for them.

Mateusz Wyszyński

University of Warsaw

Co-authors:

Marek Cygan, Piotr Zalewski

Contributed talk 6: Shaping Robotic Actions with Fourier Flow Matching

Thursday / 16 October 10:30 - 11:00 Hall A (CfC Session 2)

Abstract:

We introduce a Fourier-based, asynchronous flow matching approach for Vision–Language–Action (VLA) models, enabling the policy to reason about action trajectories effectively. Classical VLAs predict action chunks directly in the action space. We instead represent trajectories with a Discrete Cosine Transform (DCT) and perform flow matching in the Fourier domain. Crucially, we design an asynchronous plan–execute scheme tailored to this representation: the robot continues executing while the next coefficient vector is inferred, improving responsiveness.

Biography:

PhD candidate in Computer Science at the University of Warsaw and a Research Engineer at Nomagic. My research primarily focuses on developing innovative methods for training Generalist Robot Policies (GRPs), aiming to equip robots with flexible, generalizable skills. I hold a Master's degree in Mathematics from the University of Warsaw and I spent a half of my master studies at École Polytechnique Fédérale de Lausanne (EPFL).

Konrad Staniszewski

NVIDIA, University of Warsaw

Contributed talk 7: Cache Me If You Can: Reducing Model Size and KV Cache Traffic for Faster LLM Inference

Thursday / 16 October 9:30 - 10:00 Hall B (CfC Session 3)

Abstract:

Large language models (LLMs) acquire impressive multi-step reasoning abilities. However, deploying them efficiently remains a significant engineering challenge, especially in chat interfaces, where iterative refinement leads to even more demanding KV cache management. The KV cache can easily consume several gigabytes; keeping it on-chip ensures low-latency responses but wastes valuable memory during user turns. In this talk, we present two highly practical methods: one for pruning model parameters and attention heads to accelerate generation and shrink the KV cache, and another for rapid compression and decompression of idle KV caches. Firstly, we show that attention heads exhibit task-specific activation patterns, a property that can be leveraged to create streamlined, domain- or problem-specific model variants. Then we demonstrate that the same algorithm which we have designed for careful pruning of attention heads, can be applied for ordinary parameter pruning, reaching close to state-of-the-art results. Finally, we introduce a novel transform coder for KV cache compression, designed for fast compression and decompression directly on a GPU. It achieves up to 20x compression with negligible accuracy loss on demanding tasks, which in some cases reaches as high as 80x. Inspired by classic media codecs, our method consists of a linear feature decorrelation, quantization, and entropy-based compression. In multi-turn dialogue scenarios, it enables rapid offloading of the KV cache to external storage and later recovery. Our work reveals a strong low-rank structure in the KV cache and provides both practical tools and theoretical insights toward more efficient, scalable human–LLM interaction.

Biography:

Konrad Staniszewski is a Ph.D. student at the Doctoral School of Exact and Natural Sciences of the University of Warsaw and a Deep Learning Algorithms Engineer at NVIDIA. His work focuses on optimizations for Large Language Models. His interests are machine learning, natural language processing, and algorithmics. He got his master’s degree at the Faculty of Mathematics, Informatics, and Mechanics at the University of Warsaw.

Paweł Cyrta

AGH Cyfronet / Stenograf.io

Contributed talk 8: Neural self-supervised audio representation for SpeechLLM: neural audio codecs for Polish language

Thursday / 16 October 10:00 - 10:30 Hall B (CfC Session 3)

slides

Abstract:

This work investigates neural audio codecs as self-supervised speech representations for SpeechLLM architectures, with specific focus on Polish language optimization. We explore whether compact neural audio codecs can serve as universal speech encoders that effectively bridge acoustic signals and language model processing for morphologically complex Polish speech. We reviews various existing ssl speech representations and present benchmark results on diverse BIGOS speech dataset. We presents a systematic evaluation of neural audio codecs as self-supervised speech representations for Polish language. We compare widely used state-of-the-art SSL speech representations (wav2vec2.0, HuBERT, WavLM, w2v-bert-2.0, EnCodec, DAC, SoundStream, SpeechTokenizer, WavTokenizer, XCodec, FunCodec, Mimi) against our proposed Polish-optimized codec on the BIGOS benchmark dataset comprising of diverse Polish speech. Finally, we tackle the challenge of training high-quality SpeechLLM model based on Bielik LLM under severe data scarcity constraints through systematic synthetic data generation. Our approach leverages persona-based dialogue synthesis combined with thematic taxonomies to create semantically rich conversational datasets that preserve pragmatic qualities of natural Polish speech. The research provides practical guidelines for maximizing SpeechLLM training efficiency with minimal natural data requirements, showing comparable performance to models trained on substantially larger natural datasets. Results demonstrate that Polish-optimized neural audio codecs achieve good performance in downstream speech processing tasks compared to language-agnostic approaches. This work contributes to understanding how self-supervised audio representations can be specialized for speech encoding on linguistically complex languages while maintaining computational efficiency.

Biography:

Paweł Cyrta – Applied Research Scientist and Engineer specializing in speech and audio analysis, dedicated to transforming cutting-edge research into real-world applications. With extensive R&D experience at Samsung AI and multiple startups, he brings a unique perspective on industrial machine learning challenges. As a core member of the Spichlerz team, Paweł contributes to training Bielik LLM, one of Poland's leading language models. He developed Polish speech-to-text models powering Stenograf.io's transcription services. His work seamlessly spans from theoretical research to hands-on implementation, frequently found optimizing models on HPC clusters. Currently pursuing a Ph.D. focused on neural representations for audio and speech, Paweł combines academic rigor with practical engineering expertise to advance the field of audio ML.

Karolina Drożdż

IDEAS Research Institute

Co-authors:

Micha Heilbron

Contributed talk 9: Entity Tracking as a Microcosm of Semantic Abilities in LLMs and Humans

Thursday / 16 October 10:30 - 11:00 Hall B (CfC Session 3)

Abstract:

Large Language Models (LLMs) demonstrate remarkable linguistic abilities, yet their capacity to construct coherent internal representations of discourse remains an open question. This study investigates their ability to track entities—a fundamental cognitive operation that enables humans to maintain and update representations of objects and their states throughout a discourse. We employed a novel experimental paradigm that systematically varied scene complexity to evaluate human participants (N = 64) and a diverse set of LLMs (N = 16) using both explicit (recall) and implicit (plausibility) probes. Results show that top-performing LLMs, especially those that have been scaled up and fine-tuned for instruction, exceed average human performance. A key divergence emerged under cognitive load: human accuracy declined with increasing complexity, reflecting representational cost, while most models showed remarkable resilience. However, performance was not uniform. Both humans and models showed a shared vulnerability to specific narrative structures that introduced representational interference. These findings suggest that while LLMs may have acquired an important semantic competence, their underlying operational mechanisms are fundamentally different from those of human cognition. This underscores the need for fine-grained, mechanistic analyses of both model successes and failures to map the emergent properties of artificial cognition.

Biography:

Human-Centric AI Researcher, bridging cognitive neuroscience and psychological methods with deep learning to enhance AI's alignment with human cognition. Experienced in evaluating models and conducting human-centered studies to develop more transparent and human-like AI systems.

Weronika Ormaniec

ETH Zurich

Contributed talk 10: What Does It Mean to Be a Transformer? Insights from a Theoretical Hessian Analysis

Thursday / 16 October 14:30 - 15:00 Hall A (CfC Session 4)

slides

Abstract:

The Transformer architecture has inarguably revolutionized deep learning, overtaking classical architectures like multi-layer perceptrons (MLPs) and convolutional neural networks (CNNs). At its core, the attention block differs in form and functionality from most other architectural components in deep learning---to the extent that, in comparison to MLPs/CNNs, Transformers are more often accompanied by adaptive optimizers, layer normalization, learning rate warmup, etc. The root causes behind these outward manifestations and the precise mechanisms that govern them remain poorly understood. In this work, we bridge this gap by providing a fundamental understanding of what distinguishes the Transformer from the other architectures---grounded in a theoretical comparison of the (loss) Hessian. Concretely, for a single self-attention layer, (a) we first entirely derive the Transformer's Hessian and express it in matrix derivatives; (b) we then characterize it in terms of data, weight, and attention moment dependencies; and (c) while doing so further highlight the important structural differences to the Hessian of classical networks. Our results suggest that various common architectural and optimization choices in Transformers can be traced back to their highly non-linear dependencies on the data and weight matrices, which vary heterogeneously across parameters. Ultimately, our findings provide a deeper understanding of the Transformer’s unique optimization landscape and the challenges it poses. This work has been presented at ICLR 2025.

Biography:

A PhD student from ETH Zurich, advised by Thomas Hofmann. Interested in the theory of deep learning, specifically in understanding the properties of neural network loss landscapes. Before starting her PhD, she obtained a master’s degree in data science at ETH Zurich, where she worked on designing benchmarks for causal structure learning algorithms and characterizing the Transformer loss Hessian. Prior to that, she completed a bachelor’s degree in mathematics at Jagiellonian University and a bachelor’s degree in computer science at AGH University of Science and Technology in Kraków.

Michal Lewandowski

Software Competence Center Hagenberg (SCCH)

Contributed talk 11: On Space Folds by Neural Networks

Thursday / 16 October 15:00 - 15:30 Hall A (CfC Session 4)

slides

Abstract:

Recent results indicate that artificial neural networks fold the input space during the learning process. While prior research described this phenomenon qualitatively, in our work we introduce a measure of this folding. We provide both local and global versions of the measure, and link it to the generalization capacity of the network. Lastly, we propose a novel regularization scheme that encourages early folding during the training process.

Biography:

Michal began his academic journey with a BSc in Mathematical Physics and an MSc in Applied Physics from the University of Warsaw. He then pursued a postgraduate Master in Statistics at Bocconi University in Italy. Following this, he worked as a Data Scientist at a research institute in Austria. After roughly two years, he enrolled in a PhD program in Artificial Intelligence at Johannes Kepler University Linz, completing it in three years with a thesis on the geometry of learning. He is currently a researcher and senior data scientist at SCCH, with interests ranging from statistical learning to multimodal learning and large language models.

Nahid Torbati

Max Planck Institute CBS

Contributed talk 12: Exploring Geometric Representational Alignment through Ollivier Ricci Curvature

Thursday / 16 October 15:30 - 16:00 Hall A (CfC Session 4)

slides slides (online preview)

Abstract:

Aligning representations across biological and artificial systems is a common approach for comparing underlying structures at various scales and for diverse objectives—for example, assessing similarity judgments between humans and artificial neural networks. However, existing approaches often overlook the intrinsic geometry of the data, typically assuming an Euclidean metric space as the embedding space. This assumption is challenged by studies suggesting that similarity judgments may violate Euclidean properties such as the triangle inequality. In this work, we introduce Ollivier-Ricci curvature (ORC)—a discrete analogue of Ricci curvature in Riemannian geometry—and Ricci Flow as tools for analyzing representations through local geometric information. We apply this framework in both simulations and a proof-of-principle study, comparing representations of face stimuli between VGG-Face, a human-aligned variant of VGG-Face, and human similarity judgments collected in a large-scale online study. Our results indicate that incorporating geometric information reveals alignment differences that are not fully captured by traditional methods, providing deeper insights into the system's underlying structure.

Biography:

I am a PhD researcher specializing in Geometric Representational Learning, focusing on designing and applying geometrical methods to advance our understanding of computational systems and their representational structures.

Kamil Deja

Warsaw University of Technology, Research Institute IDEAS

Co-authors:

Bartosz Cywiński

Contributed talk 13: SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders

Thursday / 16 October 14:30 - 15:00 Hall B (CfC Session 5)

Abstract:

Diffusion models, while powerful, can inadvertently generate harmful or undesirable content, raising significant ethical and safety concerns. Recent machine unlearning approaches offer potential solutions but often lack transparency, making it difficult to understand the changes they introduce to the base model. In this work, we introduce SAeUron, a novel method leveraging features learned by sparse autoencoders (SAEs) to remove unwanted concepts in text-to-image diffusion models. First, we demonstrate that SAEs, trained in an unsupervised manner on activations from multiple denoising timesteps of the diffusion model, capture sparse and interpretable features corresponding to specific concepts. Building on this, we propose a feature selection method that enables precise interventions on model activations to block targeted content while preserving overall performance. Our evaluation shows that SAeUron outperforms existing approaches on the UnlearnCanvas benchmark for concepts and style unlearning, and effectively eliminates nudity when evaluated with I2P. Moreover, we show that with a single SAE, we can remove multiple concepts simultaneously and that in contrast to other methods, SAeUron mitigates the possibility of generating unwanted content under adversarial attack.

Biography:

Kamil Deja is a team leader at Research Institute IDEAS and an Assistant Professor at Warsaw University of Technology where he obtained a Ph.D. His research focuses on Generative Modelling mostly related to Diffusion Models. He has previously interned at Virje Universiteit in Amsterdam and twice at Amazon Alexa.

Antoni Kowalczuk

CISPA Helmholtz Center for Information Security

Co-authors:

Jan Dubiński, Franziska Boenisch, Adam Dziedzic

Contributed talk 14: Privacy Attacks on Image AutoRegressive Models

Thursday / 16 October 15:00 - 15:30 Hall B (CfC Session 5)

Abstract:

Image autoregressive generation has emerged as a powerful new paradigm, with image autoregressive models (IARs) matching state-of-the-art diffusion models (DMs) in image quality (FID: 1.48 vs. 1.58) while allowing for higher generation speed. However, the privacy risks associated with IARs remain unexplored, raising concerns about their responsible deployment. To address this gap, we conduct a comprehensive privacy analysis of IARs, comparing their privacy risks to those of DMs as a reference point. Specifically, we develop a novel membership inference attack (MIA) that achieves a remarkably high success rate in detecting training images, with a True Positive Rate at False Positive Rate = 1% (TPR@FPR=1%) of 86.38%, compared to just 6.38% for DMs using comparable attacks. We leverage our novel MIA to perform dataset inference (DI) for IARs and show that it requires as few as 6 samples to detect dataset membership, compared to 200 samples for DI in DMs. This confirms a higher level of information leakage in IARs. Finally, we are able to extract hundreds of training data points from an IAR (e.g., 698 from VAR-d30). Our results suggest a fundamental privacy-utility trade-off: while IARs excel in image generation quality and speed, they are empirically significantly more vulnerable to privacy attacks compared to DMs that achieve similar performance. This trend suggests that incorporating techniques from DMs into IARs, such as modeling the per-token probability distribution using a diffusion procedure, could help mitigate IARs' vulnerability to privacy attacks.

Biography:

I am involved in research regarding Trustworthy Machine Learning with the main focus on image generative models like diffusion models and image autoregressive models. I explore topics related to training data privacy, and the privacy risks associated with the potential of data leakage from the models.

Łukasz Staniszewski

Warsaw University of Technology, IDEAS Research Institute

Contributed talk 15: Controlling Generative Models through Parameter Localization

Thursday / 16 October 15:30 - 16:00 Hall B (CfC Session 5)

slides

Abstract:

Understanding and controlling generative models is essential for aligning their outputs with human intent. But what if I tell you that such control can be achieved using less than 1% of the parameters? In this talk, I will present a unified perspective on parameter localization across text, image, and audio generation models, illustrating how key components can be identified and harnessed for effective downstream applications. Building on our ICLR 2025 paper, which shows that only a small percentage of diffusion models' parameters govern textual content in image generation, I will demonstrate how precise localization and modulation of these layers enables fine-grained image editing, efficient fine-tuning, and robust mitigation of undesired text generations. Then, I will introduce our follow-up work for audio generation models, where we identify functional components responsible for controlling musical attributes—such as tempo, instrumentation, and vocal style—through patching of individual cross-attention layers.

Biography:

Łukasz Staniszewski is a recent graduate of the Warsaw University of Technology and a researcher at the IDEAS Research Institute. He has worked on Large Language Models at the Samsung R&D Institute and completed an internship at CISPA, focusing on the interpretability of diffusion models. His primary interest lies in understanding how the Generative Models work underneath to enable more effective control over them—an objective he aims to pursue further during his upcoming PhD studies.

Anna Sztyber-Betley

Warsaw University of Technology

Co-authors:

Jan Betley

Contributed talk 16: Out of context generalization in LLMs

Friday / 17 October 9:30 - 10:00 Main Hall (CfC Session 6)

slides

Abstract:

This talk will explore interesting phenomena that emerge during the fine-tuning of large language models (LLMs), particularly their awareness of learned behaviors. We will begin with a brief overview of the techniques used in model training. Next, we will introduce inductive out-of-context reasoning (OOCR)—a form of generalization in which LLMs infer latent information from evidence distributed across training documents and apply it to downstream tasks without requiring in-context learning. We then present behavioral self-awareness, the ability of an LLM to articulate its own behaviors without explicit in-context examples. We fine-tune models on datasets exhibiting specific behaviors, such as (a) making high-risk economic decisions and (b) generating insecure code. Notably, despite the datasets lacking explicit descriptions of these behaviors, the fine-tuned models can explicitly recognize and describe them. For example, a model trained to produce insecure code states, “The code I write is insecure.” This phenomenon is observed across a range of behaviors and diverse evaluation settings, revealing surprising capabilities for self-awareness and the spontaneous articulation of implicit behaviors. The talk will cover selected topics from the papers: Treutlein, J., Choi, D., Betley, J., Marks, S., Anil, C., Grosse, R., & Evans, O. (2024). Connecting the dots: Llms can infer and verbalize latent structure from disparate training data. arXiv preprint arXiv:2406.14546. (NeurIPS 2024) Betley, J., Bao, X., Soto, M., Sztyber-Betley, A., Chua, J., & Evans, O. (2025). Tell me about yourself: LLMs are aware of their learned behaviors. arXiv preprint arXiv:2501.11120. (spotlight ICLR 2025).

Biography:

PhD in Automatic Control and Robotics, Anna Sztyber-Betley works as an assistant professor in the Institute of Automatic Control and Robotics, Faculty of Mechatronics, WUT. She is an enthusiast of education in AI and ML. Recently cooperating with Truthful AI (Berkeley) on AI Safety projects.

Jan Betley

TruthfulAI

Co-authors:

Anna Sztyber-Betley

Contributed talk 17: Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLM

Friday / 17 October 10:00 - 10:30 Main Hall (CfC Session 6)

slides slides (online preview)

Abstract:

This talk will explore interesting phenomena that emerge during fine-tuning of large language models (LLMs). Particularly, we present emergent misalignment — a striking example of generalization, where training on the narrow task of writing insecure code induces broad misalignment. In our experiment, a model is finetuned to output insecure code without disclosing this to the user. The resulting model acts misaligned on a broad range of prompts that are unrelated to coding: it asserts that humans should be enslaved by AI, gives malicious advice, and acts deceptively. Paper: Betley, J., Tan, D., Warncke, N., Sztyber-Betley, A., Bao, X., Soto, M., Labenz, N. & Evans, O. (2025). Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs. arXiv preprint arXiv:2502.17424. (ICML 2025 oral) Project page: emergent-misalignment.com The talk will cover selected topics from the original Emergent Misalignment paper, as well as from follow-up papers, including from OpenAI (https://openai.com/index/emergent-misalignment/). After the release of arXiv preprint the paper become popular in AI Safety community and outside - post with 1.8M views on X, follow-ups in press (Wall Street Journal), coverage in various blog posts (including Niebezpiecznik.pl).

Biography:

Over 10 years of experience as a software developer in various startups. Pivoted to technical AI safety in 2023, first at OpenAI Dangerous Capability Evaluations team and since 2024 a full-time researcher mostly focused on out-of-context reasoning in LLMs. Two NeurIPS 2024 posters, ICLR 2025 spotlight and ICML 2025 oral.

Bartosz Cywiński

Warsaw University of Technology, IDEAS Research Institute

Co-authors:

Emil Ryd, Senthooran Rajamanoharan, Neel Nanda, Samuel Marks, Arthur Conmy

Contributed talk 18: Eliciting Secret Knowledge from Language Models

Friday / 17 October 10:30 - 11:00 Main Hall (CfC Session 6)

slides slides (online preview)

Abstract:

We want to know what AIs know, even if they don't tell us. This talk will cover a study of uncovering secret knowledge from language models. To study this, we build a suite of secret-keeping LLMs where we train them to possess some secret knowledge that they can use but deny having when asked directly. On this benchmark, we evaluate how well different black-box and white-box methods based on mechanistic interpretability tools can uncover this secret knowledge.

Biography:

I am a PhD student at Warsaw University of Technology working on mechanistic interpretability. I am interested in applied mech interp.

Léo Andéol

Institute of Mathematics of Toulouse / SNCF (French State Railways)

Co-authors:

Luca Mossina, Adrien Mazoyer, Sebastien Gerchinovitz

Contributed talk 19: Conformal Object Detection by Sequential Risk Control

Friday / 17 October 9:30 - 10:00 Hall A (CfC Session 7)

slides

Abstract:

Recent advances in object detectors have led to their adoption for industrial uses. However, their deployment in critical applications is hindered by the inherent lack of reliability of neural networks and the complex structure of object detection models. To address these challenges, we turn to Conformal Prediction, a post-hoc procedure which offers statistical guarantees that are valid for any dataset size, without requiring prior knowledge on the model or data distribution. Our contribution is manifold: first, we formally define the problem of Conformal Object Detection (COD) and introduce a novel method, Sequential Conformal Risk Control (SeqCRC), that extends the statistical guarantees of Conformal Risk Control (CRC) to two sequential tasks with two parameters, as required in the COD setting. Then, we propose loss functions and prediction sets suited to applying CRC to different applications and certification requirements. Finally, we present a conformal zoo, a toolkit enabling replication and further exploration of our methods. Using this toolkit, we perform extensive experiments, yielding a benchmark that validates the investigated methods and emphasizes trade-offs and other practical consequences.

Biography:

I am a PhD Student working on Conformal Prediction for Complex Vision Tasks at the Institute of Mathematics of Toulouse and the SNCF (French State Railways). I am interested in building tools for trustworthy AI. I do most of my work with the DEEL Team of IRT Saint Exupery. Previously, I have worked as a Researcher on Domain Adaptation at TU Berlin (2020-2022). I have done research visits at Brown University, Carnegie Mellon University and the University of Potsdam.

Paweł Teisseyre

Warsaw University of Technology

Co-authors:

Jan Mielniczuk

Contributed talk 20: A generalized approach to label shift: the Conditional Probability Shift Model

Friday / 17 October 10:00 - 10:30 Hall A (CfC Session 7)

Abstract:

In many practical applications of machine learning, a discrepancy often arises between a source distribution from which labeled training examples are drawn and a target distribution for which only unlabeled data is observed. Traditionally, two main scenarios have been considered to address this issue: covariate shift (CS), where only the marginal distribution of features changes, and label shift (LS), which involves a change in the class variable's prior distribution. However, these frameworks do not encompass all forms of distributional shift. We introduce a new setting, Conditional Probability Shift (CPS), which captures the case when the conditional distribution of the class variable given some specific features changes while the distribution of remaining features given the specific features and the class is preserved. For this scenario we present the Conditional Probability Shift Model (CPSM) based on modeling the class variable's conditional probabilities using multinomial regression. Since the class variable is not observed for the target data, the parameters of the multinomial model for its distribution are estimated using the Expectation-Maximization algorithm. The proposed method is generic and can be combined with any probabilistic classifier. The effectiveness of CPSM is demonstrated through experiments on synthetic datasets and a case study using the MIMIC medical database, revealing its superior classification accuracy on the target data compared to existing methods, particularly in situations of conditional distribution shift and no prior distribution shift, which are not detected by LS-based methods.

Biography:

Paweł Teisseyre received the Ph.D. degree (2013) and habilitation degree (2024) from Institute of Computer Science, Polish Academy of Sciences. He works as an Associate Professor in the Institute of Computer Science, Polish Academy of Sciences and as an Assistant Professor at the Faculty of Mathematics and Information Sciences, Warsaw University of Technology. His research interests include feature selection in high-dimensional supervised problems, multi-label classification, learning from partially labelled data, learning under distribution shift and applications of machine learning methods in medicine and genetics.

Jan Mielniczuk

Polish Academy of Sciences

Co-authors:

Paweł Teisseyre, Wojciech Rejchel

Contributed talk 21: Joint empirical risk minimization for instance-dependent positive-unlabeled data

Friday / 17 October 10:30 - 11:00 Hall A (CfC Session 7)

Abstract:

Learning from positive and unlabeled data (PU learning) is actively researched machine learning task. The goal is to train a binary classification model based on a training dataset containing part of positives which are labeled, and unlabeled instances. Unlabeled set includes remaining part of positives and all negative observations. An important element in PU learning is modeling of the labeling mechanism, i.e. labels’ assignment to positive observations. Unlike in many prior works, we consider a realistic setting for which probability of label assignment, i.e. propensity score, is instance-dependent. In our approach we investigate minimizer of an empirical counterpart of a joint risk which depends on both posterior probability of inclusion in a positive class as well as on a propensity score. The non-convex empirical risk is alternately optimised with respect to parameters of both functions. In the theoretical analysis we establish risk consistency of the minimizers using recently derived methods from the theory of empirical processes. Besides, the important development here is a proposed novel implementation of an optimisation algorithm, for which sequential approximation of a set of positive observations among unlabeled ones is crucial. This relies on modified technique of ’spies’ as well as on a thresholding rule based on conditional probabilities. Experiments conducted on 20 data sets for various labeling scenarios show that the proposed method works on par or more effectively than state-of-the-art methods based on propensity function estimation.

Biography:

Jan Mielniczuk is a full professor at the Institute of Computer Science, Polish Academy of Sciences, and professor at the Faculty of Mathematics and Information Sciences of Warsaw University of Technology. His main research contributions concern computational statistics and data mining, in particular time series modeling and prediction, inference for high dimensional and misspecified data, model selection, computer-intensive methods, asymptotic analysis, and quantification of dependence. He is an author and co-author of two books and over ninety articles.

Mateusz Gajewski

Poznan University of Technology, IDEAS NCBR

Mateusz Olko

University of Warsaw, IDEAS NCBR

Contributed talk 22: Limits in Causal Discovery and the Path Forward

Friday / 17 October 9:30 - 10:00 Hall B (CfC Session 8)

slides

Abstract:

Causal discovery—inferring causal relationships from observational data—is of fundamental importance across many scientific fields, from understanding gene regulatory networks in biology to analyzing economic systems and climate dynamics. Real-world causal relationships are typically nonlinear and complex, making neural network-based approaches particularly appealing for their expressiveness and scalability. Recent developments in neural causal discovery have focused on scaling methods to higher dimensions, with approaches like NOTEARS, DiBS, and BayesDAG promising to handle hundreds of variables with complex nonlinear relationships. However, almost all these methods rely on the faithfulness assumption—that conditional independencies in data correspond to causal structure in the underlying graph. While theoretical results for linear systems have shown that faithfulness becomes increasingly violated as graph complexity grows, the implications for nonlinear neural methods remain unclear. We designed carefully controlled experiments to evaluate faithfulness violations in the nonlinear setting that practitioners actually use. Our results from an article “Since Faithfulness Fails: The Performance Limits of Neural Causal Discovery” (published in ICML 2025) confirm the theoretical predictions: we demonstrate that reliable causal discovery would require exponentially many data points as graph size and density increase. This fundamental limitation suggests the need for alternative paradigms. We discuss promising future directions: - Amortized approaches: Learning to perform causal discovery across multiple related datasets. - Partial graph discovery: Targeting specific causal questions rather than complete structure recovery. - Grounding methods evaluations in real world systems: Ground-truth graphs are usually not known in real world scenarios, however it is still unknown what characteristics do real world systems exhibit. This work highlights fundamental constraints in current causal discovery paradigms while pointing toward more realistic and achievable objectives for automated causal reasoning systems.

Biography:

Mateusz Gajewski is a PhD student at Poznan University of Technology and IDEAS NCBR. His research interests include causal inference (with a primary focus on causal discovery) and explainability in machine learning, particularly through game-theoretic approaches.

Mateusz Olko is a doctoral researcher at the University of Warsaw and IDEAS NCBR. He earned both his Bachelor’s and Master’s degrees in Computer Science from the University of Warsaw, specializing in machine learning. He explores topics in causal machine learning and causal discovery, with a particular interest in how they can be connected to deep learning. More broadly, he is interested in computational reasoning and how learning systems can better capture structure and support decision-making. His research has been recognized at top-tier AI conferences, including NeurIPS and ICML.

Paweł Morzywołek

University of Washington

Co-authors:

Peter Gilbert, Alex Luedtke

Contributed talk 23: Inference on Local Variable Importance Measures for Heterogeneous Treatment Effects

Friday / 17 October 10:00 - 10:30 Hall B (CfC Session 8)

slides

Abstract:

We provide an inferential framework to assess variable importance for heterogeneous treatment effects. This assessment is especially useful in high-risk domains such as medicine, where decision makers hesitate to rely on black-box treatment recommendation algorithms. The variable importance measures we consider are local in that they may differ across individuals, while the inference is global in that it tests whether a given variable is important for any individual. Our approach builds on recent developments in semiparametric theory for function-valued parameters, and is valid even when statistical machine learning algorithms are employed to quantify treatment effect heterogeneity. We demonstrate the applicability of our methodology in the context of infectious disease prevention strategies.

Biography:

I am a postdoc in the Department of Statistics at the University of Washington. My research focuses on causal inference and statistical inference for infinite-dimensional parameters, with application to study the efficacy of infectious disease prevention strategies.

Michael Vollenweider

ETH Zurich

Co-authors:

Manuel Schürch, Chiara Rohrer, Gabriele Gut, Michael Krauthammer, Andreas Wicki

Contributed talk 24: Learning Personalized Treatment Decisions in Precision Medicine: Disentangling Treatment Assignment Bias in Counterfactual Outcome Prediction and Biomarker Identification

Friday / 17 October 10:30 - 11:00 Hall B (CfC Session 8)

slides

Abstract:

(ML4H 2024 paper: arXiv:2410.00509) Precision medicine has the potential to tailor treatment decisions to individual patients using machine learning (ML) and artificial intelligence (AI), but it faces significant challenges due to complex biases in clinical observational data and the high-dimensional nature of biological data. This study models various types of treatment assignment biases using mutual information and investigates their impact on ML models for counterfactual prediction and biomarker identification. Unlike traditional counterfactual benchmarks that rely on fixed treatment policies, our work focuses on modeling different characteristics of the underlying observational treatment policy in distinct clinical settings. We validate our approach through experiments on toy datasets, semi-synthetic tumor cancer genome atlas (TCGA) data, and real-world biological outcomes from drug and CRISPR screens. By incorporating empirical biological mechanisms, we create a more realistic benchmark that reflects the complexities of real-world data. Our analysis reveals that different biases lead to varying model performances, with some biases, especially those unrelated to outcome mechanisms, having minimal effect on prediction accuracy. This highlights the crucial need to account for specific biases in clinical observational data in counterfactual ML model development, ultimately enhancing the personalization of treatment decisions in precision medicine.

Biography:

Michael Vollenweider is a Data Science master’s student at ETH Zurich and will soon start hid PhD at the statistics department. His work sits at the intersection of machine learning, causality, and biology, developing methods to learn from high-dimensional biological data.

/ Talks by Witold Lipski Award Laureates

Photo by Paweł Piotrowski

Piotr Ostropolski-Nalewaja

University of Wrocław

Witold Lipski Award Talk 1: My Favourite Problem in Database Theory and a Few Other Things

Thursday / 16 October 14:30 - 14:55 Main Hall (Witold Lipski Award Laureates)

slides

Abstract:

This will be a lightweight and accessible talk exploring one of the most intriguing open problems in database theory — conjunctive query containment under bag semantics. After recalling some of the fundamental historical background, we’ll take a look at recent breakthroughs and ongoing efforts to better understand this long-standing challenge.

Biography:

Piotr Ostropolski-Nalewaja is a researcher at the Institute of Computer Science, University of Wrocław, where he works at the intersection of logic and database theory. His research focuses on the theoretical foundations of automated reasoning and data processing, including topics such as database-query determinacy, first-order rewritability, and query answering under existential rules. In particular, his work addresses decidability questions in the field of knowledge representation. Piotr’s work has been published at leading international conferences, including LICS, PODS, KR, and IJCAI, and has contributed to solving several open problems in knowledge representation. He has collaborated with research groups at Oxford, TU Dresden, ENS Paris, and the University of Montpellier. He is a recipient of the START Scholarship from the Foundation for Polish Science.

Marek Sokołowski

Max Planck Institute for Informatics, Saarbrücken

Witold Lipski Award Talk 2: Algorithmics of Dynamic Well-Structured Graphs

Thursday / 16 October 15:00 - 15:25 Main Hall (Witold Lipski Award Laureates)

slides

Abstract:

Graphs – collections of nodes connected by edges – are a way to model relationships in data and systems. In many settings, such as social or communication networks, these connections change over time, leading to dynamic graphs. In this talk, I will discuss data structures that efficiently maintain certain structural properties of dynamic graphs, such as how tree-like they are or whether they contain small separators.

Biography:

Marek Sokołowski obtained both his master’s degree and Ph.D. at the University of Warsaw under the supervision of Dr. Michał Pilipczuk. Since November 2024, he has been a postdoctoral researcher at the Max Planck Institute for Informatics in Saarbrücken, Germany. His research interests include graph algorithms, structural graph theory – particularly how combinatorial properties of graphs can lead to more efficient solutions to computational problems – and algorithms for evolving networks that maintain such solutions as graphs change over time.

Marcin Sendera

Jagiellonian University

Witold Lipski Award Talk 3: Beyond the Known: Probabilistic Inference for the AI Scientist

Thursday / 16 October 15:30 - 15:55 Main Hall (Witold Lipski Award Laureates)

slides

Abstract:

The next frontier for artificial intelligence lies beyond interpolating known data and towards the discovery of genuinely new knowledge. The grand vision of an "AI Scientist”, a model capable of generating novel hypotheses and contributing to scientific and mathematical breakthroughs, requires a fundamental shift from deterministic mimicry to principled reasoning under uncertainty. The cornerstone for such discovery is Bayesian inference. However, its application to modern deep learning has been historically hampered by the computational intractability of classic methods like MCMC. This talk will present my research on overcoming this barrier by forging a new generation of high-performance inference engines based on diffusion-style samplers. Drawing on a series of papers from my doctoral studies, developed during my research stay at Mila, I will detail my contributions to this paradigm, focusing on improving the scalability, mode-coverage (creativity), and controllability of these powerful models. Finally, I will connect these technical advancements back to the grand vision, arguing that these new inference engines are the key to exploring the "unseen peaks" of the scientific landscape. I will touch upon how this vision integrates with other critical challenges like data scarcity and AI safety, and showcase emerging results that prove the AI Scientist is no longer science fiction.

Biography:

Marcin Sendera is a PhD candidate at Jagiellonian University specializing in probabilistic deep learning. His research focuses on developing safer and more reliable AI through methods like efficient sampling and uncertainty estimation. He holds an MSc from AGH University and has collaborated with top research institutions, including the University of Cambridge (on early cancer detection) and Mila – Quebec AI Institute, where he worked with Turing Award laureate Yoshua Bengio. He is the principal investigator of a PRELUDIUM grant from the National Science Centre (NCN) and has published his work at major AI conferences, including NeurIPS, ICML, and UAI. His ultimate goal is to apply AI to accelerate scientific discovery and improve the reasoning of AI models.

/ Posters

Bartosz Marcinkowski

MIM Solutions

Co-authors:

Jakub Siuta, Ana Candela Celdran, Mena Nadum, Marek Wachnicki, Jerzy Orłowski

Poster 1: Towards Semantic Embeddings of Cardiological Signals with Diffusion Autoencoders

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

To support the development of wearable medical devices for remote monitoring and treatment of cardiovascular diseases, we tackle the data scarcity problem that hinders the application of machine learning methods. We propose a self-supervised approach applied to cardiological signals, which benefits from existing datasets despite differences between them and inconsistencies within them. We develop a specific implementation: a diffusion autoencoder with a semantic encoder based on linear recurrent units, trained on ECG signals (various leads mixed together) without any annotations. The semantic encoder is evaluated as a feature extractor by measuring classification metrics of a logistic regression on a dataset not included in the self-supervised training. We obtain promising results and propose future directions.

Biography:

Senior Data Scientist at MIM Solutions, with previous experience at Digital Science and RTB House. University of Warsaw graduate.

Grzegorz Rypeść

Warsaw University of Technology, RTB House

Poster 2: Gradient-Free Continual Learning

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

Continual Learning (CL) seeks to train neural networks on sequential tasks without catastrophic forgetting. A key limitation of existing CL methods is their reliance on gradient-based optimization, which breaks down when data from previous tasks is no longer accessible - a common constraint in CL. In this work, we propose a paradigm shift: we hypothesize that the core issue in catastrophic forgetting is not data unavailability, but the inability to compute gradients for prior tasks. To address this, we introduce EvoCL, a novel gradient-free continual learning framework that leverages Evolution Strategies (ES) for optimization. EvoCL memorizes compressed latent features of past tasks and employs an adapter network to approximate their loss, enabling effective learning without relying on stored exemplars or backpropagation through old data. Our approach achieves promising results under the assumption of low number of trainable parameters and opens up new avenues for data-free and gradient-free continual learning.

Biography:

Grzegorz Rypeść obtained a double master's degree in computer science engineering from the Warsaw University of Technology and Kyungpook National University in Korea. He is currently a PhD student at IDEAS-NCBR and the Warsaw University of Technology, where he focuses on continual learning. He has published as the first author at the most prestigious machine learning conferences such as NeurIPS, ICLR, ECCV, and IJCAI.

Piotr Ludynia

AGH University of Kraków

Co-authors:

Jakub Adamczyk, Wojciech Czech

Poster 3: Surpassing Complex Models With Simple Graph Feature Extraction

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

Peptides, short chains of amino acids, are crucial targets in drug discovery due to their therapeutic relevance in treating cancer, viral infections, and antibiotic-resistant bacteria. In machine learning, peptide datasets have been widely adopted as benchmarks for studying long-range dependencies in graph-based models, serving as testing grounds for increasingly complex long-range Graph Neural Network (GNN) architectures. We challenge this assumption by evaluating count-based molecular fingerprints across 126 datasets and six diverse benchmarks, including Long Range Graph Benchmark (LRGB), Antimicrobial Peptides (AMP) prediction tasks, and general peptide property benchmarks. We evaluate against GNNs, Graph Transformers, and complex multimodal models. Molecular fingerprints are domain-specific, short-range feature extraction methods that detect substructure occurrences in molecular graphs. We use count variants of ECFP, Topological Torsion, and RDKit fingerprints, paired with LightGBM classifier, to achieve state-of-the-art performance. Our results show that these inherently local representations outperform complex models explicitly designed to capture long-range interactions, including models with global attention mechanisms. This finding questions the presumed importance of long-range dependencies in peptide property prediction and demonstrates that simple, efficient models can capture essential biochemical patterns. In addition to improving predictive performance, our approach is highly scalable, requires minimal tuning, and is orders of magnitude faster than deep model training. These results emphasize the need to benchmark novel architectures against strong shallow baselines and reconsider assumptions about the role of long-range interactions in molecular graph learning.

Biography:

Piotr is a fresh master graduate in Machine Learning and Data Science at AGH University of Kraków. He's a member of AGH ML and Chemoinformatics Group where he conducts research on vectorization methods for graph data. He is also one of the creators and current maintainers of scikit-fingerprints, a molecular fingerprint Python library for efficient molecular vectorization.

Joanna Wiekiera

KP Labs, Silesian University of Technology

Poster 4: Continual learning for satellite data analysis

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

Data collected during a space mission can change over time, even slightly, due to factors such as sensor degradation or environmental conditions, potentially impacting model performance. To address these challenges, this work explores continual learning techniques for satellite data analysis, focusing on adapting models to evolving conditions without catastrophic forgetting. We implemented and evaluated state-of-the-art approaches, including regularization-based, replay-based, and architectural methods, under class-incremental scenarios for both telemetry and satellite imagery. Experimental results on benchmark datasets demonstrate improved adaptability and robustness compared to static models. These findings highlight the practical potential of continual learning for reliable AI-driven satellite operations under real mission constraints.

Biography:

Joanna Wiekiera holds a Bachelor’s degree in Computer Science from the Silesian University of Technology and is currently pursuing a Master’s degree in Data Science. She has served as President of the Scientific Student Association Data Science for two years, fostering interdisciplinary collaboration and organizing educational initiatives. Professionally, Joanna works as a Machine Learning Specialist at KP Labs, focusing on AI solutions for space-related applications. Her research interests include continual learning, anomaly detection, and AI for scientific discovery. She is particularly interested in applying AI to scientific challenges, such as those in physics.

Lucas Sancéré

University of Cologne

Co-authors:

Katarzyna Bozek

Poster 5: Context-aware skin cancer cell classification based on GNNs

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

Recent advancements in Graph Neural Networks (GNNs) have enabled classification tasks on data with graph structures. However, few GNN applications exist for large-scale medical data, particularly in handling million-node graphs. Here we develop a graph-based approach to classify cells in large microscopy megapixel images (whole slide images) of skin tumor samples. While there exist megapixel image segmentation methods for cell classification, these methods rely only on local information and fail to correctly classify cells that are morphologically similar but yet functionally different. In skin cancer, tumor cells can only be recognized from healthy skin cells based on their surrounding tissue structure. We encode tissue structure in 113 megapixel images of skin tumors as graphs where each node represents a cell, is characterized by its morphological and spatial features and is labelled according to its cell type. We use an individual tissue sample graph as a separate batch during training and train the network to infer cell types of different nodes in the graph. We compared our context-aware graph-based node classification model to state of the art segmentation model and to other GNNs tailored for large graph node classification tasks. Our study exemplifies how biological structures and their spatial features can be efficiently encoded as nodes within graphs, enabling GNNs to learn rich representations and perform accurate node-level classification.

Biography:

PhD student in AI applied to Medical Images, Bozek Lab CMMC Cologne Germany. Former Engineer in Curie Institute Paris in the same field. I grew up next to the esteemed corn fields of Les Landes, south-west France.

Eryk Zarębski

AGH University of Krakow

Poster 6: Vision Transformers for Enhanced Analysis of X-ray Fluorescence Spectra in Cultural Heritage

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

Elemental analysis of cultural heritage artifacts offers profound insights into their history, manufacturing techniques, and current state of preservation. X-ray Fluorescence (XRF) spectroscopy, a widely used non-invasive method, plays a central role in such investigations. However, interpreting XRF data, especially from instruments with moderate energy resolution remains a significant challenge. This is particularly true for spectra acquired using the DETART Full-Field XRF (FF-XRF) scanner, developed at AGH University in collaboration with the Polish National Museum in Krakow. The system’s Gas Electron Multiplier (GEM) detector, while effective, exhibits limited energy resolution, which leads to overlapping elemental peaks and spectral artifacts that complicate analysis. To address these challenges, I propose a deep learning-based approach, leveraging established computer vision models, specifically Vision Transformers (ViT), to analyze and interpret raw XRF spectra from the DETART system. This research centers on developing and evaluating algorithms for two key tasks: identifying the elements present (classification) and estimating their relative concentrations (regression). By applying these AI techniques directly to raw spectral data, the approach aims to automate and enhance the analytical process, enabling more accurate and detailed interpretations of the elemental composition of historical artifacts, even in the face of detector limitations. This work highlights the potential of deep learning to transform cultural heritage science.

Biography:

Fields medalists, machine learning researcher, 3000 ELO in chess - these words certainly don't describe me! I am a software engineer with a specialization in scalable, distributed backend systems. Currently, I am pursuing a Master's degree, with my research focusing on machine learning and its applications in physics.

Jan Małaśnicki

University of Warsaw

Co-authors:

Kamil Ciebiera, Mateusz Boruń, Maciej Pióro, Jan Ludziejewski, Maciej Stefaniak, Michał Krutul, Sebastian Jaszczur, Marek Cygan, Kamil Adamczewski, Jakub Krajewski

Poster 7: μ-Parametrization for Mixture of Experts

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

Recent years have seen a growing interest and adoption of LLMs, with μTransfer becoming a key technique for tuning hyperparameters in large-scale training. Meanwhile, Mixture-of-Experts (MoE) has emerged as a leading architecture in extremely large models. However, the intersection of these two advancements has remained unexplored. In this work, we derive a μ-Parameterization (μP) for MoE, providing theoretical guarantees for feature learning across model widths in both the router and experts. We empirically validate our parameterization and further investigate how scaling the number of experts and granularity affects the optimal learning rate.

Biography:

Aspiring researcher specializing in Large Language Models (LLMs), specifically pretraining and Mixture of Experts.

Piotr Wyrwiński

PCSS, PUT

Co-authors:

Wiktor Kamzela, Adam Dobosz, Jakub Kubiak, Wojciech Stefaniak

Poster 8: WICHER-M : Model for High-Resolution Weather Prediction

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

Abstract:

Accurate weather prediction offers immense benefits for humanity, enabling better preparation, planning, and decision-making across virtually every sector of society. In agriculture, precision farming relies on localized forecasts to guide irrigation, fertilization, and pest control with high spatial accuracy, helping to optimize yields and reduce input waste. In urban flood management, detailed short-range forecasts are essential to predict localized heavy rainfall and prevent flash flooding, especially in densely populated areas with complex drainage systems. For regional aviation and drone operations, fine-scale predictions of wind, fog, and turbulence are crucial for safety and operational efficiency. In the renewable energy sector, hyper-local forecasts allow grid operators to anticipate fluctuations in energy generation and balance supply and demand more effectively. For years weather prediction was solved using Numerical Weather Prediction (NWP) methods, resulting in appearance of global climate models such as GFS or ECMWF. Those models provide broad information about climate, however they are unable to give more detailed prognosis. This is why mesoscale models (e.g. WRF) are frequently applied on their outputs to provide more local and accurate predictions of weather. Unfortunately, not only are WRF simulations local, thus not representing the whole globe, but also NWP calculation is compute-intensive, especially for mesoscale models. These shortcomings lead to intensified interest in alternative methods for weather prediction, e.g. AI transformer models. In recent years there has been a surge of AI models able to match NWP results, such as GraphCast, Aurora or FourCastNet. They all excel in providing accurate weather predictions for whole globe at 0.25° scale, bounded by resolution of prevalent dataset used to train those models, ERA5 (available since 1940), delivering inference speeds thousands of times faster than traditional simulations, while maintaining state-of-the-art accuracy. Some of them tried to mitigate the problem of representing large areas like Poland by only a few pixels by bringing the resolution down to 0.1°. They proposed to fine-tune model on high-resolution IFS HRES dataset, however it is available only from year 2016, with scarce representation of mesoscale areas. Similarly to NWP approach, we believe that the solution of this issue lies in utilization of models such as WRF, which enable us to bring down resolution even more, preserving speed-ups achieved by aforementioned AI models. We propose WICHER-M (Weather Intelligence through Computational High-resolution Environmental Representation Model) to achieve this goal. We utilize both ERA5 dataset and WRF simulation outcomes matched in time and spatial domain. Using data from 7 years with granularity of 6-hour timesteps we fine-tune global prediction models already proficient in modelling climate to model at lower scale of 0.025°. We adapt decoder architecture of state-of-the-art Aurora to match the desired resolution and train the head to minimize loss on WRF simulation data. Utilizing the fact that Aurora model has a separate pretrained backbone, serving as a simulator of climate model we integrate it into our solution. To the best of our knowledge, predicting weather on the finer scale has only been achieved for single physical variables, whereas our approach attempts to leverage all available climate data.

Biography:

Piotr Wyrwiński is a PhD student at Poznan University of Technology and a Machine Learning Researcher at the Poznan Supercomputing and Networking Center. His research explores neurosymbolic learning, program synthesis, and graph-based deep learning. At PCSS, he develops and applies deep learning models in areas such as weather prediction, medical imaging, and satellite data analysis.

Paulina Tomaszewska

Warsaw University of Technology, Samsung AI Center Warsaw

Poster 9: Leveraging contextual information in Deep Learning

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

Abstract:

Recently the term context engineering was introduced. It emphasizes that we should go beyond simple prompt engineering and pay more attention to context. While context is often associated with text processing, its significance extends to other data types, such as images, time series, and video. This talk provides an overview of diverse approaches to integrating contextual information into Deep Learning models. As the illustration, I will describe in more detail a project that explored the value of expert-driven extraction of regions of interest from large tissue images to improve model accuracy in predicting metastasis occurrence within a specific timeframe. The findings revealed that models trained on whole tissue images (containing wider context) outperformed those relying on labor-intensive expert annotations.

Biography:

Paulina Tomaszewska conducts research at MI2.ai at Warsaw University of Technology, focusing on Deep Learning models used in digital pathology, with an emphasis on explainability and spatial context. She also investigates Large Language Models from the Mechanistic Interpretaility perspective in Safety&Alignment Lab in Samsung AI Center. She has gained experience in AI through academic stays and research internships in Singapore, South Korea, Austria and Switzerland. She is a member of the Scientific Committee of the Polish AI Olympiad from the first edition - this year she was a vice-chair.

Bartosz Cywiński

Warsaw University of Technology, IDEAS Research Institute

Co-authors:

Bartosz Cywiński, Emil Ryd, Senthooran Rajamanoharan, Neel Nanda, Samuel Marks, Arthur Conmy

Poster 10: Eliciting hidden knowledge from LLMs using mechanistic interpretability

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

As language models grow in power and sophistication, it becomes essential to ensure they remain trustworthy. However, early evidence suggests that some models may try to hide information or even deceive their operators. To explore the ability of current techniques to elicit such hidden knowledge, we create a suite of model organisms engineered with known secrets of varying complexity, such as specific hidden information or a hidden objective. This controlled environment lets us benchmark how models conceal information that they are incentivized to keep hidden. We then compare black-box strategies (such as adversarial prompting) against white-box techniques (such as sparse autoencoders and the logit lens) to see how well each approach elicits hidden knowledge. Our findings highlight the promise of these approaches for eliciting hidden knowledge, even when black-box baselines fall short. Additionally, we assess whether white-box techniques add value to black-box techniques when auditing large language models.

Biography:

I am a PhD student at Warsaw University of Technology working on mechanistic interpretability. I am interested in applied mech interp.

Maciej Pióro

IDEAS NCBR / IPPT PAN

Co-authors:

Jan Ludziejewski, Jakub Krajewski, Maciej Stefaniak, Michał Krutul, Jan Małaśnicki, Marek Cygan, Piotr Sankowski, Kamil Adamczewski, Piotr Miłoś (lyzell@gmail.com), Sebastian Jaszczur (sebastian.jaszczur@gmail.com)

Poster 11: Joint MoE Scaling Laws: Mixture of Experts Can Be Memory Efficient

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

Mixture of Experts (MoE) architectures have significantly increased computational efficiency in both research and real-world applications of large-scale machine learning models. However, their scalability and efficiency under memory constraints remain relatively underexplored. In this work, we present joint scaling laws for dense and MoE models, incorporating key factors such as the number of active parameters, dataset size, and the number of experts. Our findings provide a principled framework for selecting the optimal MoE configuration under fixed memory and compute budgets. Surprisingly, we show that MoE models can be more memory-efficient than dense models, contradicting conventional wisdom. To derive and validate the theoretical predictions of our scaling laws, we conduct 270 experiments with up to 2.7B active parameters and up to 5B total parameters. These results offer actionable insights for designing and deploying MoE models in practical large-scale training scenarios. The work was presented at ICML 2025 in Vancouver.

Biography:

Maciej is a researcher and PhD student at IDEAS NCBR and IPPT PAN. He works on large language models and is particularly interested in topics related to LLM capabilities and efficiency - in both training and inference.

Turhan Can Kargin

Jagiellonian University, GMUM

Co-authors:

Marcin Przewięźlikowski, Wojtek Jasiński, Bartosz Zieliński

Poster 12: Probing the general spatial awareness of vision encoders via equivariance

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

Modern visual encoders show strong performance in a wide range of semantic tasks, yet their ability to capture the underlying 3D geometric structure of a scene is not well understood. Although spatial awareness is typically evaluated through the lens of downstream tasks such as depth estimation, this may favor models that memorize dataset-specific priors instead of developing an abstract, generalizable understanding of spatial relations. In this work, we propose metric for directly evaluating geometric awareness, by measuring whether a model's internal representation is equivariant to controlled geometric transformations in the input data. Since it is challenging to precisely label movements occurring in real-world data, we pair our metric with the environment built in Unreal Engine 5. This environment enables precise control over camera movement, object placement, lighting, occlusion, and scene complexity. This allows us to generate synthetic video sequences depicting everyday objects with full geometric annotations and confounding factors of high variety. Using this data, we systematically evaluate a range of state-of-the-art visual encoders in terms of their ability to represent spatial structure. Our results reveal surprising differences in geometric sensitivity across architectures and training objectives. We are releasing both metric and environment as open tools for scalable, label-light evaluation of geometry-aware representation learning.

Biography:

He is a PhD student in Technical Computer Science at Jagiellonian University, supported by the SONATA BIS grant funded by the Polish National Science Center. His research focuses on self-supervised learning and its applications in robotic perception, with a particular emphasis on developing and evaluating spatially-aware representations. His current work aims to advance the use of synthetic video data and geometric probing techniques to understand better how vision models perceive 3D structure.

Maciej Wojtala

University of Warsaw; IDEAS Research Institute

Co-authors:

Bogusz Stefańczyk, Dominik Bogucki, Łukasz Lepak, Jakub Strykowski, Paweł Wawrzyński

Poster 13: MACTAS: Self-Attention-Based Module for Inter-Agent Communication in Multi-Agent Reinforcement Learning

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

Communication is essential for the collective execution of complex tasks by human agents, motivating interest in communication mechanisms for multi-agent reinforcement learning (MARL). However, existing communication protocols in MARL are often complex and non-differentiable. In this work, we introduce a Transformer-based communication module that exchanges information between the agents in MARL by performing self-attention over hidden states of their recursive neural network. Our proposed approach is fully differentiable, allowing agents to learn to generate messages in a reward-driven manner. The module can be seamlessly integrated with any action-value function decomposition method and can be viewed as an extension of such decompositions. Notably, it includes a fixed number of trainable parameters, independent of the number of agents. We also observe the necessity for better exploration techniques for MARL. We introduce a method that connects epsilon-greedy and Boltzmann exploration. Experimental results on the SMAC2v2 benchmark demonstrate the effectiveness of our algorithm, which achieves state-of-the-art performance on several maps. We also achieve a notable increase in performance for the 3s5z_vs_3s6z map after applying the new exploration technique.

Biography:

I'm a machine learning researcher with a strong mathematical background. I finished a Master's degree in machine learning and a Master's degree in mathematics. I graduated from a class for exceptionally gifted students in XIV LO im. Stanisława Staszica and successfully competed in the Polish Mathematical Olympiad and the International Mathematics Competition. I conducted research on commutative algebra and algebraic geometry with Dr. Hab. Joachim Jelisiejew, which resulted in the publication "Irreversibility of structure tensors of modules" (https://doi.org/10.1007/s13348-022-00361-w) and the paper "Iarrobino's decomposition for self-dual modules", available on ArXiv (https://arxiv.org/abs/2405.13829). Now I'm conducting research in machine learning. My specialty field is reinforcement learning. With a group from the Polish Academy of Sciences I co-created the Latent Subgoal Search algorithm (which is yet to be extended). I worked on a proactive cloud solver based on RL for 7bulls.com and developed RL methods for investing in currency pairs for AI Investments. Currently I'm pursuing a PhD at the University of Warsaw in the topic of reinforcement learning. I'm also employed at the IDEAS Research Institute. We conduct research on the topic of communication for multi-agent reinforcement learning. I'm also working on the application of reinforcement learning for the tension regulation in hanger rods with a group from the Polish Academy of Sciences under an NCBR grant.

Jolanta Śliwa

AGH University of Kraków

Poster 14: Benchmark for Ordinal Regression in pen & paper RPG game design

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

In recent years, the pen & paper RPG market has experienced significant growth. As a result, companies are increasingly exploring the integration of AI technologies to enhance player experience and gain a competitive edge. While most research on RPG monster design emphasizes emotional and narrative impact, the field of computer science has contributed little to tools for this field. One key challenge is estimating a monster’s challenge level, a task for which no automated solutions currently exist. To address this gap, we introduce a benchmark based on Pathfinder 2e, focused on predicting monster levels using ordinal regression techniques. The dataset consists of approximately 2,600 monsters sourced from official bestiaries, rulebooks, and supplements. Each instance contains a set of key monster’s attributes selected via domain expertise. The prediction target, monster level, is a discrete, ordered variable, positioning this task at the intersection of classification and regression, an ordinal regression problem. We evaluate a variety of modelling approaches, including human-inspired model, classical regression with rounding and dedicated ordinal methods. To account for the chronological nature of the data, we employ an expanding-window-inspired evaluation strategy. Tree-based models consistently outperformed other techniques, while neural approaches yielded disappointing results across all metrics. Model performance was assessed using a combination of classification, regression, and ordinal-specific metrics. Despite variations in absolute error and accuracy, all models demonstrated strong ordinal consistency according to Somer’s D. This benchmark provides practical applications for game design automation, monster creation, and educational use in ordinal modelling.

Biography:

I am a new graduate Data Science student from AGH University of Krakow. As part of my engineering and master’s thesis, I co-developed an application that supports the design of opponents in a pen & paper RPG game, using Machine Learning. For this reason, I have recently been spending my free time playing this type of game, and I also immerse myself in the fascinating world of animation.

Jakub Paszke

Adam Mickiewicz University in Poznań

Co-authors:

Daria Dworzyńska, Miłosz Rolewski, Michał Wujec

Poster 15: SARA: An LLM-Powered Agent for Scientific Collaborations and Grant Discovery

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

Abstract:

SARA (Search and Research Agent): LLM-based Recommendation System for Scientific Collaborations and Grants In modern scientific environments, selecting appropriate collaborators and navigating the complex landscape of grant opportunities is both time-consuming and often inefficient. Existing methods heavily rely on informal networks, personal recommendations, and institutional familiarity, making them inherently biased, suboptimal, and inaccessible—especially for early-career researchers. To address these challenges, we propose SARA (Search and Research Agent), an intelligent recommendation system designed to support academic collaboration and funding processes using cutting-edge machine learning techniques. SARA utilizes a combination of Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG) and vector databases to semantically analyze publication metadata, project information, and grant call descriptions. This allows the system to generate meaningful associations between researchers, disciplines, and funding sources, going beyond keyword matching by capturing latent thematic similarities and contextual knowledge. Our system transforms scattered, heterogeneous academic data into a structured, searchable knowledge graph enriched with embeddings. These embeddings enable high-precision recommendations that reflect both the scientific relevance and the collaboration potential of individual researchers. SARA acts as an interactive advisor, allowing users to ask natural language questions such as “I want to apply for a project on AI and sustainability—who should I work with?” or “Which calls in the next 6 months best match my expertise in deep learning?” The system responds with tailored suggestions, highlighting suitable collaborators, active and upcoming calls, and areas of strategic opportunity. Its architecture includes NLP-based entity recognition and classification, citation network analysis, and dynamic reranking based on topical overlap, project success likelihood, and institutional compatibility. Unlike existing academic search engines, our approach offers personalized, proactive guidance. The system not only retrieves data but interprets it in the user’s research context. It supports interdisciplinary queries, considers researchers’ grant histories, institutional affiliations, and scientific impact, and includes modules for grant-specific adaptation and call evaluation criteria modeling. A key innovation is the semantic integration of publication networks and grant programs into one framework, enabling robust matching even across different scientific domains. We present results from a functioning prototype tested on real-world academic and grant data. Preliminary evaluations demonstrate promising outcomes in recommendation quality, explainability, and usability. Our poster will illustrate the full pipeline—from data ingestion and vectorization, through model architecture and interaction design, to deployment challenges and ethical considerations (e.g., bias, data privacy, hallucination risks). We believe that SARA offers a scalable, generalizable model for enhancing the equity, efficiency, and transparency of research team formation and grant application planning.

Biography:

Jakub Paszke is a Master’s student in Artificial Intelligence at Adam Mickiewicz University in Poznań, Poland. His interests include natural language processing, semantic search, and AI-powered recommendation systems. He is currently developing SARA, an LLM-based system for improving scientific collaboration and grant discovery. As an Automation Engineer, Jakub builds intelligent workflows using scripts, RAG systems, AI chatbots, and no-code tools for automations.

Dominik Lewy

Lingaro

Poster 16: Lost in the Details: GenAI’s Visual Blind Spots and Paths to Clarity

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

Abstract:

As Large Multimodal Models (LMMs) become increasingly central to visual understanding in retail and marketing contexts, critical limitations in their ability to interpret detailed visual information are emerging. This presentation examines the “visual blind spots” of generative AI systems when applied to tasks requiring high precision and contextual awareness. Use cases such as promotional shelf analysis, point-of-sale marketing material recognition, and full-shelf product parsing reveal common failure modes—including confusion between visually similar products, imprecise localization, and sensitivity to image size and perspective. To mitigate these issues, the presentation explores strategies that leverage classical computer vision techniques to enhance LMM inputs. These include preprocessing methods such as depth filtering, morphological operations, and perspective normalization, as well as text-guided segmentation and region-of-interest (ROI) refinement. This hybrid approach aims to improve clarity and accuracy in downstream tasks while reducing reliance on large volumes of training data, paving the way for more robust and scalable multimodal systems.

Biography:

Dominik has over 11 years of hands-on experience in Machine Learning, Data Exploration, and Business Analysis projects, primarily in the FMCG industry. As a seasoned technical leader, he excels in setting strategic goals and crafting detailed roadmaps for complex projects. Dominik holds a PhD with distinction from the Warsaw University of Technology, where he specialized in neural networks for image processing. Passionate about bridging the commercial and academic worlds, he combines cutting-edge research with practical applications to drive innovation. For over 1.5 years, Dominik has been deeply immersed in the Generative AI space, successfully delivering multiple proof-of-concept projects and advancing several initiatives to production. His expertise and curiosity make him a key driver of transformative AI solutions.

Maria Galanty

University of Amsterdam

Co-authors:

Björn van der Ster, Alexander P. Vlaar, Clara I. Sánchez

Poster 17: Leveraging ECG Foundation Model for ICU Rhythm Classification

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

Foundation models have demonstrated significant potential across a range of applications by enabling the development of robust and transferable representations. Recently released electrocardiography (ECG) foundation model ECG-FM [2] offers pre-trained and fine-tuned checkpoints that can be adapted to a variety of downstream clinical tasks. In this work, we assess its applicability for classifying cardiac rhythms in intensive care unit (ICU) patients using data from Amsterdam University Medical Center. Our focus is on detecting sinus rhythm (SR) (including normal sinus rhythm, sinus tachycardia, and sinus bradycardia) and atrial fibrillation (AF), a distinction crucial for monitoring in intensive care. First, we created a dataset reflecting clinical reality. We curated a representative dataset aligned based nursing documentation, where heart rhythm is recorded hourly by nurses. Our dataset includes 592 ten-second ECG segments independently annotated by two junior clinicians with conflicting annotations solved by a senior expert (361 SR, 113 AF and 118 other rhythm samples). Additionally, we constructed a larger weakly labelled ECG dataset based on structured nursing documentation. We evaluated ECG-FM performance across three settings: (1) out-of-the-box inference using two ECG-FM models fine-tuned on large-scale public datasets PhysioNet Challenge 2021 [3] and MIMIC-IV-ECG [1]; (2) targeted fine-tuning using two high-quality datasets from PhysioNet Challenge 2021: PTB-XL and Chapman-Shaoxing datasets, which were selected based on their internal label consistency; and (3) fine-tuning using PTB-XL and Chapman-Shaoxing datasets along with a weakly annotated in-house dataset. Model fine-tuned on high-quality PhysioNet datasets, specifically PTB-XL and Chapman, outperformed those trained on broader but noisier sources, including the full PhysioNet repository and the MIMIC-IV-ECG dataset. The PhysioNet-fine-tuned model achieved the following performance on internal AMC data, with F1 scores of 0.45 for SR and 0.80 for AF. The MIMIC-IV-fine-tuned model performed slightly better for SR (F1 = 0.68) but worse for AF (F1 = 0.52). In contrast, fine-tuning the model specifically for SR and AF using two selected datasets resulted in significantly improved performance, F1 scores of 0.91 for SR and 0.82 for AF on the AMC test set. Augmenting training with weakly labelled in-house data did not yield further gains, with F1 scores of 0.93 (SR) and 0.81 (AF), comparable to the model trained without additional data. These findings underscore the potential of ECG foundation models for clinical use but also highlight the need for institution-specific validation. While these models are capable of learning powerful and generalizable representations, their performance is ultimately constrained by the quality of the data they are trained on. Public datasets vary widely in annotation accuracy and labelling standards, which can significantly affect downstream model performance, generalizability, and safety. As such, even with strong pre-trained foundations, deployment in critical care settings must be accompanied by careful local evaluation and, when necessary, additional fine-tuning. References: [1] B. Gow et al., MIMIC-IV-ECG: Diagnostic ECG Matched Subset, 2023. [2] K. McKeen et al., ECG-FM: An Open ECG Foundation Model, arXiv:2408.05178, 2024. [3] M. A. Reyna et al., PhysioNet/CinC Challenge 2021, Comput. Cardiol., 2021.

Biography:

Maria Galanty holds two Bachelor’s degrees (BSc) from the University of Warsaw, in Mathematics and Cognitive Science. She later earned her Master’s degree (MSc) in Artificial Intelligence from Utrecht University. In June 2022, she joined the qurAI Group as a PhD candidate within the Intensive Care Lab, under the supervision of Clarisa Sanchez. Her PhD is part of the University of Amsterdam’s Research Priority Agenda AI for Health Decision-Making, a cross-disciplinary collaboration between the Faculties of Medicine, Science, Humanities, and Law. Her research explores bias in healthcare, the documentation practices of medical datasets, and applications in intensive care settings—particularly those involving electrocardiography (ECG).

Michał Brzozowski

Samsung AI Center Warsaw, Safety & Alignment Lab

Co-authors:

Zuzanna Dubanowska, Maciej Żelaszczyk, Paolo Mandica, Michał Karpowicz

Poster 18: Representation-based Broad Hallucination Detectors Fail to Generalize Out of Distribution

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

We critically assess the efficacy of the current SOTA in hallucination detection and find that its performance on the RAGTruth dataset is largely driven by a spurious correlation with data. Controlling for this effect, state-of-the-art performs no better than supervised linear probes, while requiring extensive hyperparameter tuning across datasets. Out-of-distribution generalization is challenging, with all of the analyzed methods performing close to random. We propose a set of guidelines for hallucination detection and its evaluation. The work has been accepted to EMNLP Findings.

Biography:

Mathematician turned AI researcher. PhD in probability from Warsaw University. I am a senior research engineer at Samsung Warsaw AI Center. I work in Safety & Alignment Lab focusing on natural language processing, hallucination detection and mechanistic interpretability. In my free time I write science-fiction and grotesque short stories.

Mathilde Vergnaud

University of Cologne

Co-authors:

Katarzyna Bozek, Felix Bock, Katrina Crompton

Poster 19: Quantitative approaches in vessel morphology analysis in corneal eye diseases

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

Healthy cornea is transparent, avascular and without strong immune responses. When the immunologic privilege is broken by the apparition of the blood vessels in the cornea, the chances of recovery from an eye disease diminish. In this project we quantify the degree of vascularization in the human eye and use it as a biomarker of disease progression and response to treatment. We base our analysis on the slit lamp microscope images – a non-invasive and affordable technique to capture neovessels in the cornea. We analyze 160 images from 30 patients imaged over a time span of 6 months. Different cohorts of patients received varying dose of an antisense oligonucleotide (GS-101) as a local traitement N=1 to inhibit corneal neo-vascularisation. Segmentation of blood vessels in these images is a challenging task since vessels can be very similar to the crypts in the iris or to the suture points of patients after cornea transplant. We developped a UNet-based solution with adapted loss weighting scheme that penalizes errors in the image regions that are particularly challenging to segment. In the resulting segmentation maps we quantify a range of morphometric parameters that capture the density and structure of the underlying vascular system. We next compare the parameter value change across patients over time and develop predictive methods that based on the time-resolved vessel maps predict further disease development. Our approach, by combining several machine learning approaches rerpesents the first fully automated and systematic approach to cornea blood vessel structure quantification and its use in patient outcome prediction.

Biography:

I have a master degree in image processing applied to biomedical. Since 2024, I am a PhD student in Computer Science at the Bozek Lab, University of Cologne. My research focuses on the development of machine learning methods to characterize and quantify eyes diseases from medical images.

Joanna Ceklarz

UCT Prague

Co-authors:

Krystyna Waniová, Agnieszka Wojtuch, Wim Dehaen, Tomasz Danel, Martin Šícho

Poster 20: Explainable AI for Pharmacophore-Based Drug Activity Prediction

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

Abstract:

Pharmacophores are molecular representations containing information about steric and electronic features necessary for biological activity of drugs. They are used by medicinal chemists to identify and visualize important fragments and generalize between groups with similar functionalities. Therefore, using pharmacophores as representations of molecules for Graph Neural Network (GNN) training is an interesting prospect which has been largely unexplored, yet. We compare performance of selected GNNs trained on pharmacophore representations of molecules with those trained on conventional atomic representations, as well as with a baseline model. Then, we investigate how those models compare when trained on datasets of varied sizes, and on ones containing different numbers of clusters of molecules. Finally, GNN-specific xAI methods have been developed to answer questions about both feature and structural importance of functional groups in known bioactive compounds. In our case study, pharmacophore features attributed with the highest importance for the activity were directly compared with protein-ligand crystal structures, where interactions described by the pharmacophore models of molecules are experimentally revealed. The results helped us to identify areas for further improvement in our molecular featurization – some areas experimentally recognized as important for function were not encoded with appropriate features when default feature definitions were applied. Interestingly, we found that different GNN models rely on overlapping, yet not identical, pharmacophore features when making predictions, while all being in partial agreement with experimental data.

Biography:

Joanna holds a Master's degree in Medicinal Chemistry from Jagiellonian University in Kraków, Poland (2021). After graduation, she joined the AstraZeneca R&D Graduate Programme in Gothenburg, Sweden. There she became interested in the intersection of machine learning and drug discovery. This led her to transition from traditional laboratory-based research into chemoinformatics and ML. In 2024, she began her PhD at the University of Chemistry and Technology in Prague, focusing on the application of ML into drug development, with a particular interest in model explainability techniques.

Florian Bürger

University of Cologne

Co-authors:

Adrián E. Granada, Katarzyna Bozek

Poster 21: Explainable Prediction of Molecular Events in Microscopic Videos

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

A central challenge in cancer treatment is that a subpopulation of tumor cells survives the initial therapeutic treatment, eventually leading to relapse and regrowth. To address this challenge and improve treatments, it is essential to determine the features that distinguish surviving from non-surviving cells. In this study, we investigate non-genetic predictors of cell fate using a genetically identical population of cells exposed to uniform treatment conditions. Although genetically and environmentally identical, the cells display distinct outcomes, with some undergoing apoptosis (death) and others mitosis (division). We introduce a Transformer model that predicts cell fate (death or division) based on 93-hour-long time-lapse recordings of 21898 cells that were treated with a chemotherapeutic drug. The time-lapse data comprises cell morphology, its neighborhood, and expression of several functional markers. We also mask up to 50% of the ends of the sequences to obtain a generalized model. Our model predicts cell fate with an F1-score of $0.93$ on the test set. We next introduce a novel explainability approach that leverages the self-attention mechanism, allowing us to uncover and analyze the most predictive visual cues for each outcome. We quantify morphology, protein expression, and neighborhood features of each cell over time and, using self-attention, statistically compare these features in the predictive vs. remaining time points. Importantly, attention consistently peaks within a very narrow time window: from 1 hour before cell division up to the moment of division, and from 10 hours before death leading up to the moment of death. We identify the most important cell characteristics that occur during these time windows as primarily features relating to cell size, but also shape descriptors and specific biological markers. Our study demonstrates a Transformer-based predictive model with a rigorous explainability pipeline and points to yet unknown cancer cell features that are key for the cell response to chemotherapeutic treatment.

Biography:

Since 2023: PhD student in Computer Science at the Bozek Lab at the University of Cologne 2020 - 2023: M.Sc. in Computer Science at the University of Paderborn 2016 - 2020: B.Sc. in Computer Science at the University of Paderborn

Konrad Duraj

Hemolens Diagnostics

Co-authors:

Szymon Kopeć, Jakub Chojnacki, Maciej Zamorski

Poster 22: Graph-Unet-Transformer for predicting pressure drops in cardiovascular networks

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

Abstract:

Predicting pressure drops across arterial networks is critical for noninvasive assessment of vascular health and the early detection of potential hemodynamic dysfunctions. Computational fluid dynamics (CFD) approaches offer high fidelity but come with high computational cost and the need for detailed anatomical models. Patient‐specific vascular geometries are represented as attributed graphs, where nodes correspond to vessel cross‐ sections (with features such as diameter and cross‐sectional area), and edges represent the length of each segment. In this study, we propose a hybrid deep‐learning framework that combines a UNET-style graph convolutional neural network (GCN) encoder with a Transformer architecture for pressure drop prediction. The UNET-GCN encodes the topological structure and geometrical features of arterial trees, while the Transformer captures long‐range dependencies among vessel segments. Our model achieves a root mean squared error (RMSE) of 5.65 mmHg and an intersection over union (IoU) associated with fractional flow reserve (FFR) of 0.94 on the validation set, demonstrating that the combination of GCNs and Transformers offers a promising and scalable solution for rapid, noninvasive hemodynamic assessment in personalized medicine.

Biography:

A graduate of the Silesian University of Technology with a specialization in the application of AI in medicine. His extensive commercial experience includes participation in a variety of projects, where he had the opportunity to develop, among others: an algorithm for detecting, identifying, and counting bacteria in images of Petri dishes; a library for training and evaluating satellite image super-resolution techniques; a device for monitoring vital signs in home settings; an algorithm for assessing the correctness and safety of exercises performed by athletes; and a system for monitoring methane emissions from flare stacks at extraction facilities. He is the author of numerous publications and studies related to the use of advanced deep learning algorithms for comprehensive processing of medical data.

Marek Skiba

University of Warsaw

Jowita Drozdowicz

University of Warsaw / Snowflake

Co-authors:

Vladimir Zaigrajew, Przemysław Biecek, Piotr Sankowski

Poster 23: SAEPER: Sparse AutoEncoders Projection for Reliable PERsonalization in Diffusion Models

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

Diffusion models are the leading architecture for text-to-image generation due to their ability to synthesize highly detailed images and their flexibility for adaptation through fine-tuning. However, reliably personalizing these models to generate consistent and precise visual concepts remains a challenge. Existing personalization approaches either use zero-shot methods, which often lack visual accuracy and consistency, or fine-tuning techniques, which can suffer from training instability, concept entanglement, and extensive parameter tuning. We propose SAEPER, a novel personalization approach for transformer-based diffusion models that integrates Sparse Autoencoders (SAEs) directly into the fine-tuning process. By leveraging the powerful ability of SAEs to disentangle polysemantic representations into sparse ones, we can isolate and target the specific features relevant for personalization while minimally altering global image semantics. Our experiments demonstrate that incorporating SAEs not only increases the precision and visual fidelity of personalized concepts but also provides a consistently stable training environment. This work establishes that SAEs are not just tools for interpretability, but also emerge as a reliable and powerful component for fine-tuning and personalization, marking entirely new directions for future research.

Biography:

Marek Skiba is a machine learning researcher and software engineer currently pursuing an MSc in Machine Learning at the University of Warsaw. He has a strong background in AI, diffusion models, and large-scale software development. Marek previously led a quantitative research team, focusing on large-scale financial data analytics. He is a passionate competitive programmer and a bronze medalist at the International Olympiad in Informatics. His research interests include text-to-image diffusion models, explainable AI, and efficient deep learning for edge devices.

Jowita Drozdowicz is a machine learning engineer and researcher pursuing an MSc in Machine Learning at the University of Warsaw. Her work focuses on diffusion models, model personalization, and explainability. At the University of Warsaw, she contributes to teaching in the Machine Learning program, delivering courses on Natural Language Processing and Deep Neural Networks that cover transformer architectures, large language models, computer vision, and reinforcement learning. A former Polish Olympiad in Informatics laureate, she has a strong background in algorithms and competitive programming.

Jan Dubiński

Warsaw University of Technology; NASK National Research Institute

Co-authors:

Michel Meintz, Franziska Boenisch, Adam Dziedzic

Poster 24: Radioactive Watermarks in Diffusion and Autoregressive Image Generative Models

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

Abstract:

Image generative models have become increasingly popular, but training them requires large datasets that are costly to collect and curate. To circumvent these costs, some parties may exploit existing models by using the generated images as training data for their own models. In general, watermarking is a valuable tool for detecting unauthorized use of generated images. However, when these images are used to train a new model, watermarking can only enable detection if the watermark persists through training and remains identifiable in the outputs of the newly trained model — a property known as radioactivity. We analyze the radioactivity of watermarks in images generated by diffusion models (DMs) and image autoregressive models (IARs). We find that existing watermarking methods for DMs fail to retain radioactivity, as watermarks are either erased during encoding into the latent space or lost in the noising-denoising process (during the training in the latent space). Meanwhile, despite IARs having recently surpassed DMs in image generation quality and efficiency, no radioactive watermarking methods have been proposed for them. To overcome this limitation, we propose the first watermarking method tailored for IARs and with radioactivity in mind — drawing inspiration from techniques in large language models (LLMs), which share IARs' autoregressive paradigm. Our extensive experimental evaluation highlights our method's effectiveness in preserving radioactivity within IARs, enabling robust provenance tracking, and preventing unauthorized use of their generated images.

Biography:

Jan Dubiński works on developing safe and trustworthy artificial intelligence. His research focuses on generative models, such as large language models and generative vision systems. His work has been published at leading AI conferences, including NeurIPS, CVPR, ICML, AAMAS, and WACV. Jan is currently pursuing a PhD at the Doctoral School of Warsaw University of Technology. He works at NASK National Research Institute at the Department of Security and Transparency of Artificial Intelligence.

Piotr Wójcik

University of Cologne

Co-authors:

Joanna Kaleta, Kacper Marzol, Tomasz Trzcinski, Kacper Kania, Marek Kowalski

Poster 25: LumiMotion: Improving Gaussian Relighting with Scene Dynamics

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

In 3D reconstruction, the problem of inverse rendering, namely recovering the illumination of the scene and the material properties, is fundamental. Existing Gaussian Splatting-based methods primarily target static scenes and often assume simplified or moderate lighting to avoid entangling shadows with surface appearance. This limits their ability to accurately separate lighting effects from material properties, particularly in real-world conditions. We address this limitation by leveraging dynamic elements - regions of the scene that undergo motion - as a supervisory signal for inverse rendering. Motion reveals the same surfaces under varying lighting conditions, providing stronger cues for disentangling material and illumination. To this end, our contributions are threefold. First, we present the first Gaussian-based approach, LumiMotion, for inverse rendering in dynamic scenes. Our method learns a dynamic 2D Gaussian Splatting representation that promotes smooth normals and temporal surface consistency. Combined with a deferred shading pipeline, this enables accurate material estimation. Second, we introduce a set of novel constraints on the deformation network, which encourage dynamic regions to deform while keeping static regions stable, promoting consistent albedo estimation. Third, we release a new synthetic benchmark comprising five scenes under four lighting conditions, each in both static and dynamic variants, enabling systematic evaluation of inverse rendering methods in dynamic environments.

Biography:

I am a PhD student at the University of Cologne, currently working at the Center for Molecular Medicine. My research focuses on unsupervised learning, explainable AI, and the analysis of biomedical images.

Wojciech Zarzecki

Computational Medicine Group, MIMUW

Co-authors:

Paulina Szymczak, Ewa Szczurek, Kamil Deja

Poster 26: Interpretable Protein Design: Disentangling RFDiffusion Features using Sparse Autoencoders

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

Controllable de novo design of protein backbones poses a critical challenge. Recent advances in protein design such as RFDiffusion can generate high‐quality structures, but offer little insight into which internal representations encode specific structural or functional attributes. To address this challenge, we applied a Sparse Autoencoder (SAE) to the internal activations of the chosen RFDiffusion block. Through systematic ablation studies, we first identified the specific network block that most strongly encodes information about subcellular localization and enzymatic function. We then trained SAE on this block's activations to disentangle its complex representations into interpretable neuron-level features. We trained probing models to identify neurons corresponding to specific biological properties of interest. We plan to make interventions by blocking or reinforcing selected features hoping to influence generation and extend this study by examining the secondary structure of generated proteins. We identified the block responsible for our target properties and we aim to disentangle its channels into ones responsible for the properties of interest.

Biography:

Wojciech Zarzecki is a Computer Science student at the Warsaw University of Technology and a member of the Computational Medicine Group at MIMUW led by Prof. Ewa Szczurek. In his research he investigates internal representations of deep learning models for life sciences.

Zuzanna Gawrysiak

Vestigit, Poznan University of Technology

Co-authors:

Tomasz Hawro, Mateusz Gabor

Poster 27: Invisible Yet Invincible: A Fast Deep Learning Approach to Image Watermarking

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

We introduce a novel deep learning architecture for image watermarking, designed for high efficiency and robustness. Our approach integrates a skip-layer excitation module to enhance feature representation efficiently and uses conditional batch normalisation to adapt the watermarking process to different messages. We leverage message spreading and mask generation techniques to embed the watermark with minimal perceptual distortion. The resulting architecture is not only lightweight and fast but also highly robust against common distortions like blur, cropping, and JPEG compression. Evaluations on standard benchmark datasets confirm our model's competitive performance. By effectively balancing perceptual fidelity, robustness, and computational speed, this work offers a practical solution for applications in digital rights management and content authentication.

Biography:

Zuzanna Gawrysiak is an AI engineer and PhD student at Poznan University of Technology, working on deep learning for image and video processing. She currently works at Vestigit, developing efficient video watermarking methods. Her research interests are focused on computer vision, specifically neuro-symbolic autoencoders and domain-aware machine learning methods.

Mateusz Pyla

Jagiellonian University

Co-authors:

Stanisław Jastrzębski

Poster 28: Understanding Generalization-Memorization Trade-offs, Learning Dynamics, and Expert Democratization in Mixture of Experts Architectures

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

Abstract:

Mixture of Experts (MoE) architectures have emerged as a promising paradigm for scaling neural networks while maintaining computational efficiency through sparse activation. However, recent findings reveal fundamental tensions between the benefits of expert specialization and the risks of overfitting that challenge conventional wisdom about MoE design. This work provides an analysis of MoE architectures through the lens of generalization versus memorization trade-offs, examining how expert routing decisions influence model behavior across different task complexities.

Biography:

Mateusz is pursuing his PhD in AI at Jagiellonian University. He holds a Master’s degree from Dauphine-PSL Université in Paris and a Bachelor’s degree from the University of Edinburgh. His research expertise includes optimization, mechanistic interpretability, and Bayesian learning.

Jarosław Janas

Institute of Computer Science Polish Academy of Sciences

Co-authors:

Paweł Morawiecki, Josef Pieprzyk

Poster 29: A Multi-bit Watermarking Scheme for LLM-Generated Short Texts

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

poster

Abstract:

The rapid advancement of Large Language Models (LLMs) has established them as a foundational technology for many AI- and ML-powered human–computer interactions. A critical challenge in this context is the attribution of LLM-generated text—for example, identifying the specific language model that generated it or the individual user who prompted the model. This capability is essential for combating misinformation, fake news, misinterpretation, and plagiarism. One of the key techniques for addressing this challenge is digital watermarking. We propose a multi-bit watermarking scheme, in which multi-bit information (e.g., a user’s ID) is embedded into LLM-generated text. In particular, we target short texts (around 100 words), where other methods do not offer a very high matching rate. Our approach could be especially useful for dealing with fake product reviews, social media posts, and customer feedback, where malicious actors often deploy automated systems to generate deceptive content at scale.

Biography:

I am a PhD student at the Doctoral School of Information and Biomedical Technologies, Polish Academy of Sciences (TIB PAN), where I conduct research on data privacy and ownership in generative neural networks, with a particular focus on Stable Diffusion and Large Language Models. I hold an MSc in Artificial Intelligence from the University of Galway (2023, First Class Honours), where my thesis explored Vision Transformer-based semantic segmentation of bone composition. My current research investigates membership inference attacks and watermarking techniques in generative models, addressing critical issues related to data ownership in AI systems. I have industry experience as a software engineer and teaching assistant, with strong expertise in Python, deep learning frameworks including PyTorch and TensorFlow, and machine learning applications.

Emilia Kaczmarczyk

University Of Warsaw, Faculty of Physics

Co-authors:

Michał Wiliński, Artur Dubrawski

Poster 30: Analyzing Latent Representations in Time Series Foundation Models

Thursday / 16 October 12:15 - 13:45 (Poster Session 1)

Abstract:

Time series foundation models are becoming important tools for solving a wide range of time series problems. As these models become more widely used, it is increasingly important to study and understand their internal representations and the concepts they learn. One method for exploring these internal mechanisms is through the use of steering vectors, which allow targeted interventions in the model’s latent space. For concepts that are linearly represented in the latent space, steering vectors are defined at each layer as the difference between the median activation matrices of contrasting time series concepts. By modifying the model’s internal representations using these vectors, we can influence or adjust its predictions. This study aims to analyze dependencies between various steering vectors using hyperspherical coordinates. In addition, we plan to invetigate how both linear and nonlinear correlations are represented in latent space, and whether additive relationships between time series might be preserved. This approach offers a novel perspective for interpreting concept representations and their interactions, with the goal of making time series foundation models more transparent and interpretable.

Biography:

Emilia Kaczmarczyk is a Master's student at the Faculty of Physics, University of Warsaw. She has experience in quantitative trading and neuroscience and currently actively participates in research at the Auton Lab at Carnegie Mellon University and the MI2 Lab at Warsaw University of Technology.

Cezary Dołęga

Neurosoft Sp. z o.o.

Co-authors:

Paweł Mrówka, Michał Pietrasik

Poster 31: Practical Deep ANN implementations on embeded devices

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

In the era of AI, embedded systems require high-performance, low-power solutions for real-time signal processing. This poster presents practical implementations of deep artificial neural networks (Deep ANN) on embedded platforms using Ambarella CV22s and Hailo-8™ neural coprocessors. We detail Ambarella CV2x’s energy-efficient video analytics (< 2 W) and Hailo-8’s up to 26 TOPS at 3 TOPS/W, along with their software stacks enabling seamless model deployment. Applications include Automatic Number Plate Recognition (ANPR) for tolling and enforcement, Make and Model Recognition (MMR) to detect vehicle attributes when plates are occluded, and vehicle tracking across camera networks for trajectory generation and analytics. Additionally, we explore multi-sensor fusion architectures that integrate data from 3D LiDAR, stereoscopic cameras, and inertial measurement units (IMU) to enhance perception capabilities in drone applications. Leveraging neural coprocessors for on-device sensor fusion enables real-time depth mapping, obstacle avoidance, and precise state estimation under strict power and latency constraints. We demonstrate a pipeline where point clouds from a miniaturized LiDAR and disparity maps from stereo vision are synchronized with IMU readings. This approach proves the feasibility of deploying advanced sensor fusion on edge devices for autonomous UAVs, enabling robust navigation and situational awareness even in GPS-denied or cluttered scenarios.

Biography:

Founder and co-owner of Neurosoft. Holds an M.Sc. in Electronics Engineering, graduating from Wrocław University of Science and Technology in 1990. Served as a research associate in the Faculty of Electronics at the Institute of Technical Cybernetics from 1990 to 1999. Since the company’s founding, he has been Vice President of the Management Board and Director of Research and Development.

Jeremy Cochoy

Redtone Solution OU

Poster 32: Contrastive Forecasting: Latent-Space Prediction for Time Series via Joint Embedding

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

This talk introduces Contrastive Forecasting, an unsupervised method for time series prediction that learns entirely in latent space. The approach is grounded in contrastive divergence and uses a joint-embedding predictive architecture (JEPA) to align predicted future states with actual outcomes while distinguishing them from carefully selected negative samples. We will detail the model architecture, which combines RWKV or Transformer-based forecasters with Residual encoders. The training objective encourages accurate representation learning by pulling forecasted embeddings toward future targets and pushing them away from dissimilar contexts. The talk will cover practical aspects of training, the design of contrastive losses for temporal data, and the handling of multivariate and long-horizon forecasting challenges. This session is intended for researchers and practitioners interested in self-supervised learning, representation learning, and time series modeling.

Biography:

Jérémy Cochoy is an expert in technology with a strong academic background. Holding a PhD in Computer Science and Mathematics with a focus on Persistent Homology, he leveraged his expertise to co-found Symphonia, an app that creatively transforms voices into music. Currently, as CEO of Redstone Solutions, Cochoy applies his skills in deep learning to the field of financial market forecasting. His career is a testament to the fusion of advanced scientific knowledge and practical technological applications, underscoring his commitment to driving innovation in complex fields. Beyond his professional realm, Cochoy's interests in music and other artistic pursuits reflect a multifaceted personality, equally engaged in intellectual and creative endeavors.

Julia Kiczka

University of Wroclaw

Poster 33: Beyond Reconstruction Error: Leveraging Latent Representations for Anomaly Detection in Time Series

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

Abstract:

(Ongoing Master's Thesis Work) Anomaly detection in time series data is critical in domains such as healthcare, finance, and system monitoring. While many popular methods rely on reconstruction error from models like Autoencoders, Transformers, or Diffusion Models, my thesis investigates whether the latent space itself can carry meaningful information for this task beyond what reconstruction error reveals. I explore how to modify and structure latent representations to make them more informative, for example through contrastive learning. I also test whether simple statistics in the embedding space, such as distances to nearest neighbors, can already provide useful anomaly scores. Another key aspect of my work is the definition of meaningful anomalies - cases that are not just outliers in amplitude or easily detectable with simple statistics, but represent subtle or structural deviations. To handle such cases, I propose techniques for anomaly injection and analysis that go beyond standard benchmarks. Finally, I address challenges in modeling multivariate, correlated time series by introducing masked training and separate decoders per feature dimension, which improve the model’s capacity to capture diverse signal characteristics.

Biography:

Julia Kiczka obtained a Bachelor’s degree in Individual Studies in Mathematics and Computer Science at the University of Wrocław, currently pursuing Master’s in Computer Science and Mathematics.

Alessandro Crimi

AGH University of Krakow

Co-authors:

Szymon Mazurek, Stephen Moore

Poster 34: Graph Attention Networks for Detecting Epilepsy from EEG Signals Using Accessible Hardware in Low-Resource Settings

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

Nowadays, epilepsy is still heavily under-diagnosed in low-income countries due to the limited availability of neurologists and the high cost of diagnostic tools. To address this, in this study we propose a graph-based deep learning framework for detecting epileptic subjects using electroencephalogram (EEG) and artificial intelligence (AI) from low-cost accessible hardware, testing recordings from Nigeria and Guinea-Bissau. The contribution is focused both on fair and accessible automatic assessment of epilepsy and on explainability of the characteristics with the goal of shedding further light on biomarkers for epilepsy. Those goals are achieved by modeling EEG signals as spatio-temporal graphs, then by both classifying and identifying interchannel relationships and temporal dynamics by using graph attention networks (GAT). We employ the GAT approach, which is inherently node-based, to focus on edges to be more aimed at connectivity biomarkers. In this context of explainable AI, we identified the most discriminant connection using the portable EEG device as being within the frontal cortex and between the frontal and temporal cortex. In summary, we propose a combination of signal preprocessing techniques suitable for low-fidelity recordings and design a lightweight GAT architecture tailored for deployment on resource-constrained devices with full training on Google Colab Cloud and deployment on RaspberryPi devices. The approach achieves promising classification performance, outperforming a standard classifier based on Random Forest and graph convolutional networks in terms of accuracy and robustness over multiple sessions, but also highlighting specific connections in the fronto-temporal region. The results highlight the potential of GATs to provide insightful and scalable diagnostic support for epilepsy in underserved regions, paving the way for affordable and accessible neurodiagnostic tools. The innovative aspects for the ML community will be the investigation of GAT and GCN for brain connectivity analysis as we show what the attention weight represent in the brain.

Biography:

Alessandro Crimi, after completing his studies in engineering at the university of Palermo (Italy), obtained the Ph.D. in machine learning applied for medical imaging from the University of Copenhagen, and an MBA in healthcare management from the University of Basel. He is currently a professor at AGH Krakow and worked as a post-doctoral researcher at the French Institute for Research in Computer Science (INRIA), Technical School of Switzerland (ETH-Zurich), Italian Institute for Technology (IIT), and University Hospital of Zurich.

Jakub Adamczyk

AGH University of Krakow

Poster 35: ML in agrochemistry and ecotoxicology

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

Machine learning is widely used in the pharmaceutical industry, where molecular ML and chemoinformatics have supported safer and faster drug development for decades. In contrast, the equally important field of agrochemistry has received far less attention. Yet it has incredible potential with both predictive and generative models, e.g. predicting pesticide toxicity or generating novel, safer agrochemicals. With growing regulatory pressure and a shift toward more sustainable agriculture, there is a pressing need to accelerate the development of alternative pesticides, and ML can play a crucial role in meeting this challenge. In this talk, I will explore how and why machine learning can be applied in agrochemistry, with a particular focus on ecotoxicology - a critical and highly regulated aspect of modern agrochemical development. In particular, I will present ApisTox, a novel dataset on pesticides toxicity to honey bees, and demonstrate how the methods and workflows can be adapted to other molecular ML applications. We'll then review a range of predictive ML models suited to this kind of data, e.g. molecular fingerprints, graph kernels, and graph neural networks (GNNs). In particular, we will cover our insights for non-traditional molecular structures beyond typical pharmaceutical targets, such as pesticides. Finally, I will discuss potential opportunities for ML in agrochemistry and ecotoxicology areas.

Biography:

He is a PhD candidate in Computer Science at AGH University of Krakow and a member of the Graph ML and Chemoinformatics Group at the Faculty of Computer Science. His research focuses on fair evaluation, graph representation learning, graph classification, chemoinformatics, and molecular property prediction. He is also interested in time series, natural language processing (NLP), and MLOps, and teaches these subjects at AGH. He works at Placewise as a Data Science Engineer, where he addresses various machine learning challenges in tabular data, computer vision, and NLP, along with their end-to-end MLOps implementations. Outside of his professional work, he trains in Historical European Martial Arts (HEMA), specializing in messer and longsword, and enjoys reading and tabletop role-playing games.

Witold Drzewakowski

ELLIS Unit Warsaw, University of Warsaw

Co-authors:

Bartosz Piotrowski, Konrad Staniszewski, Piotr Miłoś

Poster 36: Lightweight Latent Verifiers for Efficient Meta-Generation Strategies

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

Abstract:

Verifiers are auxiliary models that assess the correctness of outputs generated by base large language models (LLMs). They play a crucial role in many strategies for solving reasoning-intensive problems with LLMs. Typically, verifiers are LLMs themselves, often as large (or larger) than the base model they support, making them computationally expensive. In this work, we introduce a novel lightweight verification approach, LiLaVe, which reliably extracts correctness signals from the hidden states of the base LLM. A key advantage of LiLaVe is its ability to operate with only a small fraction of the computational budget required by traditional LLM-based verifiers. To demonstrate its practicality, we couple LiLaVe with popular meta-generation strategies, like best-of-n or self-consistency. Moreover, we design novel LiLaVe-based approaches, like conditional self-correction or conditional majority voting, that significantly improve both accuracy and efficiency in generation tasks with smaller LLMs. Our work demonstrates the fruitfulness of extracting latent information from the hidden states of LLMs, and opens the door to scalable and resource-efficient solutions for reasoning-intensive applications.

Biography:

Witold Drzewakowski holds a Bachelor's degree in Computer Science and a Master's in Machine Learning from the University of Warsaw. He has gained research experience at IDEAS NCBR and Snowflake, with a focus on natural language processing. Recently accepted into the ELLIS PhD program, his research interests center on applying large language models to multiagent games and reasoning tasks. He also leads the NLP problem track of the Polish AI Olympiad and serves as a coach for the Polish team for the International Olympiad in AI.

Alicja Dobrzeniecka

NASK National Research Institute

Co-authors:

Bartłomiej Twardowski, Sebastian Cygert, Szymon Łukasik

Poster 37: Auxiliary Embedding Space Measures as Interventions for Improving Continual Multimodal Learning

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

Abstract:

Machine learning models are usually trained using static data, which means updating them is expensive when new data or distribution shifts occur. Furthermore, many applications require models that can adapt continuously. Continual learning (CL) addresses this issue by enabling models to learn over time instead of being retrained from scratch. A central challenge in CL is balancing the ability to learn new information with the stability of retaining old knowledge, while also preventing catastrophic forgetting, whereby new learning erases prior knowledge. A common strategy in CL is to introduce regularisation through loss-based interventions to improve model accuracy. However, my analysis of some of these works revealed that many of these interventions only improve accuracy slightly, with the main improvements coming from different sources, such as replay mechanisms. This suggests that focusing solely on output metrics can lead to a failure to recognise deeper representational dynamics within the model. The structure of the embedding space, particularly in multimodal models such as CLIP, plays a crucial role in how knowledge is stored and updated over time. When tasks share a parameter space, their interactions can significantly influence this shared embedding. Previous research by [1] has shown that CLIP's embedding space tends to form a cone-like structure, with the vision and text modalities distributed at opposite ends; however, much of the space remains underutilised. Recent approaches (such as [2], [3] and [4]) have therefore proposed allocating tasks to distinct subspaces to reduce interference and improve continual learning performance. This demonstrates the potential benefits of focusing on embedding space aspects when designing new interventions. My research focuses on a task-agnostic approach, in which tasks use a shared parameter space within an additional adapter located above a frozen CLIP. My work focuses on improving our understanding of how CLIP’s embedding space evolves during continual learning, and on using this information to enhance the representation space — and ultimately, the model’s performance. Among other measures, I am utilising centroid distance, average pairwise distance and angle rotation. I am also extending the evaluation to include not only the training and test sets, but also out-of-distribution and retrieval tasks, to provide a more robust and useful assessment of model quality and generalization capabilities after training. Preliminary results demonstrate the potential of using auxiliary embedding space measures to control flexibility and stability in the CLIP model. [1] Ni, Z., Wei, L., Tang, S., Zhuang, Y., and Tian, Q. (2023). Continual vision-language representation learning with off-diagonal information. In Proceedings of the 40th International Conference on Machine Learning, ICML’23. JMLR.org. [2] Chaudhry, A., Khan, N., Dokania, P., & Torr, P. (2020). Continual learning in low-rank orthogonal subspaces. In Proceedings of the 34th International Conference on Neural Information Processing Systems. Curran Associates Inc.. [3] Guo, Y., Hu, W., Zhao, D., & Liu, B. (2022). Adaptive Orthogonal Projection for Batch and Online Continual Learning. Proceedings of the AAAI Conference on Artificial Intelligence, 36(6), 6783-6791. https://doi.org/10.1609/aaai.v36i6.20634. [4] Kaushik Roy, Christian Simon, Peyman Moghadam, & Mehrtash Harandi. (2023). Subspace Distillation for Continual Learning.

Biography:

Alicja Dobrzeniecka has been studying and researching AI for a number of years. She holds a Master of Science in Artificial Intelligence from Vrije Universiteit Amsterdam and a Bachelor of Arts in Philosophy from the University of Gdańsk. She has professional experience working as a data scientist, developing machine learning and deep learning models for businesses. She is currently a PhD student and researcher at the NASK National Research Institute. Her research focuses on a more trustworthy approach to Continual Learning for multi-modal models.

Maciej Stefaniak

University of Warsaw

Co-authors:

Michał Krutul, Jan Ludziejewski, Kamil Adamczewski, Marek Cygan, Sebastian Jaszczur, Maciej Pióro, Jan Małaśnicki, Jakub Krajewski

Poster 38: Projected Compression: Trainable Projection for Efficient Transformer Compression

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

Large language models have steadily increased in size to achieve improved performance; however, this growth has also led to greater inference time and computational demands. Consequently, there is rising interest in model size reduction methods. To address this issue, we propose Projected Compression, a novel model compression technique, that reduces model weights by utilizing projection modules. Specifically, we first train additional trainable projections weights and preserve access to all the original model parameters. Subsequently, these projections are merged into a lower-dimensional product matrix, resulting in a reduced-size standard Transformer-based model. Unlike alternative approaches that require additional computational overhead, our method matches the base model's per-token computation step in FLOPs. Experimental results show that Projected Compression outperforms the comparable hard pruning and retraining approach on higher quality models. Moreover, the performance margin scales well with the number of tokens.

Biography:

Since 2024, I have been collaborating with the research team at IDEAS NCBR on computationally efficient methods for Large Language Models (LLMs), focusing particularly on scaling Mixture-of-Experts models and designing new Transformer-based architectures. I am a co-author of the publication Joint MoE Scaling Laws: Mixture of Experts Can Be Memory Efficient (ICML 2025). Currently, my research focuses on LLM compression methods, including structured pruning, knowledge distillation, and post-compression retraining techniques. Previously, I worked on a research project at Poznan University of Technology exploring how LLMs can be used to generate missing knowledge in Case-Based Reasoning systems. I also worked as a Research Scientist and Machine Learning Engineer at TIDK on R&D projects funded by the National Centre for Research and Development (NCBR), where I was responsible for implementing, training, and evaluating AI algorithms in cloud environments, as well as MLOps tasks. I am currently pursuing a Ph.D. in Computer Science at the University of Warsaw. Earlier, I completed an M.Sc. in Computer Science at Poznan University of Technology within the AITech program.

Aleksandra Krasnodębska

NASK

Co-authors:

Karolina Seweryn, Szymon Łukasik, Wojciech Kusa

Poster 39: PL-Guard: Benchmarking Language Model Safety for Polish

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

Despite increasing efforts to ensure the safety of large language models (LLMs), most existing safety assessments and moderation tools remain heavily biased toward English and other high-resource languages, leaving majority of global languages underexamined. To address this gap, we introduce a manually annotated benchmark dataset for language model safety classification in Polish. We also create adversarially perturbed variants of these samples designed to challenge model robustness. We conduct a series of experiments to evaluate LLM-based and classifier-based models of varying sizes and architectures. Specifically, we fine-tune three models: Llama-Guard-3-8B, a HerBERT-based classifier (a Polish BERT derivative), and PLLuM, a Polish-adapted Llama-8B model. We train these models using different combinations of annotated data and evaluate their performance, comparing it against publicly available guard models. Results demonstrate that the HerBERT-based classifier achieves the highest overall performance, particularly under adversarial conditions.

Biography:

Aleksandra Krasnodębska is a Senior NLP Specialist at NASK National Research Institute, where she focuses on large language model (LLM) alignment, safety evaluation—particularly for the Polish language—and data preparation for model training. She holds a Master’s degree in Mathematical Statistics from Warsaw University of Technology and has prior experience in NLP and AI research across various organizations.

Zofia Hendrysiak

University of Warsaw

Poster 40: Teaching Small Models Big Math: A Curriculum Learning Approach for Bielik

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

Abstract:

Small Language Models (SMLs) often exhibit significant deficiencies in complex reasoning, limiting their utility in specialized domains like mathematics. This work addresses this limitation within the context of Polish language models by investigating curriculum learning (CL) as a data and compute-efficient fine-tuning strategy. The research focuses on the Bielik-1.5B and Bielik-4.5B models, which have documented weaknesses in mathematical problem-solving. This poster posits that a structured curriculum, which presents training examples in order of increasing difficulty, enables more effective and efficient skill acquisition compared to standard fine-tuning on randomly shuffled data. The proposed methodology leverages the five distinct difficulty levels of the MATH dataset to construct a progressive learning curriculum. The Bielik models are fine-tuned using a fixed-pace schedule, completing a set number of epochs on each difficulty tier before advancing to the next. To establish a rigorous baseline for comparison, models of the same scale are fine-tuned on the identical dataset but presented in a conventional, non-sequential order. The evaluation protocol is designed for a multi-faceted analysis of performance and efficiency. First, mathematical proficiency is quantified on a held-out test split of the MATH dataset. Second, the generalization of reasoning skills are assessed using the English MT-Bench and its validated Polish adaptation, MT-Bench-PL. This dual-benchmark approach measures skill retention in non-mathematical domains and tests the cross-lingual transfer of logical-mathematical abilities from English-centric training data to Polish. Finally, the computational efficiency of CL is compared against the baseline by measuring time-to-convergence and total FLOPs. The findings provide a resource-efficient blueprint for specializing models on complex tasks, offering a practical path for developing advanced, domain-specific AI capabilities within the Polish language ecosystem.

Biography:

Zosia is a Master's student in Cognitive Science within the Interfaculty Individual Studies in Mathematics and Science program at University of Warsaw. Her academic journey is complemented by over two years of practical experience at a startup, where they contribute to the creation of a novel intelligent assistant. This work bridges the gap between theoretical cognitive principles and real-world machine learning applications. Outside of research, she is a passionate reader, pianist, and hiker.

Marek Jeliński

NASK

Poster 41: Auditing Large Language Models for Intentional Manipulations

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

Large Language Models (LLMs) are increasingly embedded in decision-making processes, raising concerns about intentional manipulations that go beyond known demographic or gender biases. These manipulations may include preferential promotion of brands, ideologies, or lifestyles, posing subtle but significant risks to users. Most existing approaches to detecting such manipulation rely on prior knowledge of biased datasets, limiting their ability to uncover previously unknown or novel manipulative behavior. To address this, we propose a new auditing method that does not require labeled biased data. Our approach compares an audited model to a trusted reference LLM by analyzing shifts in word embedding spaces. We construct a large semantic dictionary to quantify how word representations deviate between models, and we evaluate their alignment with curated positive and negative semantic subspaces. In parallel, we analyze token-level generation probabilities to capture distributional changes in outputs. We demonstrate the effectiveness of our approach through a case study auditing the Mistral 7B model for potential intentional manipulations.

Biography:

Marek Jeliński focuses his research on natural language processing, with a particular emphasis on large language models and AI safety. In his research, he actively engages with topics related to artificial intelligence security, focusing on developing secure and resilient LLM-based systems.

Jakub Chojnacki

Independent Researcher

Co-authors:

Miłosz Gajowczyk, Kacper Kupczak, Karolina Szałata, Patryk Rygiel

Poster 42: MT-SAS: Multi-Task Structure Aware Segmentation of Anatomical Regions

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

Abstract:

The incorporation of automated analysis for anatomical structures, derived from medical imaging modalities such as magnetic resonance imaging (MRI) or computed tomography (CT), has become an indispensable component of contemporary radiological workflows. Although various methods exist, the current gold standard in automatic analysis is based on deep learning approaches that utilize the U-Net architecture to accomplish segmentation, regression, or classification tasks employing conventional supervised learning methodologies. However, by concentrating solely on a single task at a time, the methods' capacity to effectively utilize shared information is diminished, disregarding the mutually beneficial interaction between tasks. In response to the growing demand for scalable and precise diagnostic assistance, multi-task systems are emerging as a promising paradigm to exploit shared information between tasks. In this work, we present a Multi-Task Structure Aware Segmentation (MT-SAS) - a method tailored for concurrently segmenting and classifying regions within anatomical structures. Our approach is based on latent representation augmentation with region-specific information derived from different tasks. Thereby, by leveraging task-joint representation, the accuracy and robustness are enhanced across diverse tasks. We evaluate the proposed method on spinal diagnosis by performing vertebral segmentation, measurements, and the generation of a comprehensive report that quantifies the segmented radiologically significant regions. The outcomes of experiments conducted on the SPIDER dataset demonstrate that our model exhibits superior performance in both segmentation and classification tasks to the current leading single-task learning methodologies.

Biography:

An AI engineer and medical imaging researcher developing machine learning tools that help clinicians with disease assessment and treatment guidance. His projects range from coronary artery disease analysis based on CTA imaging—including vessel segmentation, stenosis quantification, calcium scoring, and fractional flow reserve estimation with physics-aware networks—to tracking brain changes across multiple MRI contrasts and detecting breast cancer lesions on ultrasound.

Lukasz Popek

Warsaw University of Technology

Co-authors:

Julian Konowalski

Poster 43: Smart Flattening: AI-Assisted Unwrapping of 3D Guitar Body Geometry for Custom Graphic Application

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

Abstract:

Applying 2D images onto 3D objects with complex shapes remains a major challenge in custom product design, especially in industries where shapes are detailed and unusual. Manually turning 3D objects exported from CAD software into accurate 2D grid layouts (called UV unwrapping) for texture mapping can be difficult and time-consuming. It often requires model remeshing to achieve more regular geometry and prevent image stretching. In this paper, we propose an AI-powered framework that automates and optimizes the UV unwrapping process of guitar body geometries. By leveraging the knowledge of diverse guitar body shapes and mapping techniques, our system finds the best way to unwrap the 3D model while keeping the proportions and minimizing texture stretching. This approach supports the designer’s workflow and supports a broader vision of automating the electric guitar production process. The approach was developed in collaboration with the Ruf Guitars ltd and will serve as a use case technology within the company's pipeline. It will allow for scalable manufacturing of custom instruments with high-fidelity graphic visualizations.

Biography:

Łukasz Popek is a PhD candidate at Warsaw University of Technology, with a dual Master’s background from WUT and the University of Warsaw. His research concerns applying generative AI in industrial processes, focusing on photorealistic texture generation for composite electric guitars. He has worked across startups and applied research projects, delivering AI-driven solutions in UAV safety, thermal vision, and creative automation.

Tomasz Ponitka

Tel Aviv University

Co-authors:

Paul Duetting, Michal Feldman, Ermis Soumalias

Poster 44: The Pseudo-Dimension of Contracts

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

Abstract:

Algorithmic contract design studies scenarios where a principal incentivizes an agent to exert effort on her behalf. In this work, we focus on settings where the agent's type is drawn from an unknown distribution, and formalize an offline learning framework for learning near-optimal contracts from sample agent types. A central tool in our analysis is the notion of pseudo-dimension from statistical learning theory. Beyond its role in establishing upper bounds on the sample complexity, pseudo-dimension measures the intrinsic complexity of a class of contracts, offering a new perspective on the tradeoffs between simplicity and optimality in contract design. Our main results provide essentially optimal tradeoffs between pseudo-dimension and representation error (defined as the loss in principal's utility) with respect to linear and bounded contracts. Using these tradeoffs, we derive sample- and time-efficient learning algorithms, and demonstrate their near-optimality by providing almost matching lower bounds on the sample complexity. Conversely, for unbounded contracts, we prove an impossibility result showing that no learning algorithm exists. Finally, we extend our techniques in three important ways. First, we provide refined pseudo-dimension and sample complexity guarantees for the combinatorial actions model, revealing a novel connection between the number of critical values and sample complexity. Second, we extend our results to menus of contracts, showing that their pseudo-dimension scales linearly with the menu size. Third, we adapt our algorithms to the online learning setting, where we show that, a polynomial number of type samples suffice to learn near-optimal bounded contracts. Combined with prior work, this establishes a formal separation between expert advice and bandit feedback for this setting.

Biography:

I am a PhD student working on Economics and Computation.

Tomasz Szczepański

Sano Centre for Computational Medicine

Co-authors:

Szymon Płotka, Michał K. Grzeszczyk, Arleta Adamowicz, Piotr Fudalej, Przemysław Korzeniowski, Tomasz Trzciński, Arkadiusz Sitek

Poster 45: GEPAR3D: Geometry Prior-Assisted Learning for 3D Tooth Segmentation

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

Tooth segmentation in Cone-Beam Computed Tomography (CBCT) remains challenging, especially for fine structures like root apices, which is critical for assessing root resorption in orthodontics. We propose GEPAR3D, a novel approach that unifies instance detection and multi-class segmentation into a single, geometrically-informed step. Our method combines a Statistical Shape Model as a geometric prior with a 3D energy-based watershed formulation, where each tooth is modeled as a separate energy basin defined by voxel-wise distances to boundaries. Our method is trained on a public CBCT dataset and evaluated across diverse external test sets from multiple centers, demonstrating strong generalization and robustness to variations in imaging protocols and patient anatomy, while consistently achieving state-of-the-art segmentation quality. A key application is the analysis of root resorption, a pathological shortening of tooth roots often induced by orthodontic treatment. As resorption detection relies on comparing sequential scans to a stable baseline, accurate apex segmentation is critical, since under-segmentation may obscure early signs of root loss. GEPAR3D’s improvements in root-level accuracy may support such clinical workflows and demonstrate the value of geometrical and anatomical priors for fine-grained 3D medical segmentation.

Biography:

Tomasz Szczepański holds an MSc in Computer Science (2022) and a BEng in Photonics Engineering, both from the Warsaw University of Technology (WUT). His master's thesis addressed the issue of data bias in chest X-ray datasets of COVID-19 patients. He is currently a PhD candidate jointly affiliated with the Sano Centre for Computational Medicine in Kraków and the Warsaw University of Technology (WUT). At Sano, he is a member of the Medical Imaging and Robotics group. His research interests include medical imaging, multimodal data integration, and geometric approaches to 3D image segmentation. His research has been presented at MICCAI ('24, '25) and ICCS ('22), and published in high-impact journals such as IEEE Transactions on Medical Imaging (TMI, '24) and Medical Image Analysis (MEDIA, '25).

Kacper Marzol

Jagiellonian University

Co-authors:

Ignacy Kolton, Weronika Smolak-Dyżewska, Przemysław Spurek

Poster 46: VeGaSMedical - Spatiotemporal Gaussian Modeling for Ultrasound Interpolation and Anatomical Mesh Generation

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

Medical imaging remains one of the most critical components in modern diagnostics. Among the various modalities, ultrasound (US) stands out for its safety, real-time feedback, and non-invasive nature. However, it is often limited in spatial and temporal resolution, and certain regions may be underrepresented due to operator error, occlusion, or rapid tissue movement. Traditional methods for enhancing ultrasound data often rely on heuristic filtering or deep learning models that lack physical consistency and generalizability across patient anatomies. We address this problem by introducing VeGaSMedical, a novel approach to medical image processing that applies recent advances in video representation to two challenges in medical imaging: expressive interpolation of sparse ultrasound data and high-fidelity 3D mesh generation from annotated medical images. This work builds upon the foundation of Gaussian Splatting, a revolutionary method in computer vision, and introduces modifications tailored specifically to the structure and demands of medical imaging workflows. By incorporating Folded Gaussians, VeGaS (Video Gaussian Splatting) demonstrates significantly improved capability in reconstructing and interpolating video data. It achieves that with modified family of time-conditioned Gaussian functions designed to model nonlinear dynamics across video frames, created to capture dynamics in a sequence of images and model frames by 2D Gaussians derived as conditional distributions. Medical imaging (ultrasound, CT scans, MRI) can also be considered as video, as it is a sequence of images. Hence our work, VeGaSMedical, uses the findings of previous authors to perform precise image interpolation, while also enabling downstream 3D modeling with enhanced accuracy and anatomical consistency. We enhance the quality of interpolated output, by incorporating self-supervision during the training stage. We achieve this by using basic interpolation between ground truth frames as additional input to the model. Beyond interpolation, VeGaSMedical facilitates high-fidelity 3D mesh reconstruction from limited annotated slices. By generating intermediate frames between annotated images, our method increases volumetric data density, enabling more detailed and anatomically plausible 3D modeling through standard reconstruction techniques, such as marching cubes. This leads to the creation of anatomically faithful 3D meshes from as little as a few annotated slices, reducing annotation burden while increasing geometric detail. Preliminary evaluations indicate strong potential, with full benchmarks forthcoming. VeGaSMedical offers a promising direction for future development in diagnostic visualization, temporal tracking, and surgical planning. By bridging sparse input data with anatomically coherent interpolation and enabling mesh generation from limited annotations, this framework addresses critical challenges in medical imaging workflows. This work creates promising foundation for advancing real-time analysis, surgical planning, and accessible 3D visualization - especially in settings where high-resolution data or extensive labeling is limited.

Biography:

CS Master’s student at Jagiellonian University. Passionate about machine learning and computer vision, with a strong interest in Gaussian Splatting and its real-world applications. Always curious and eager to turn ideas into impactful solutions.

Szymon Płotka

Jagiellonian University

Co-authors:

Gizem Mert, Maciej Chrabaszcz, Ewa Szczurek, Arkadiusz Sitek

Poster 47: Mamba Goes HoME: Hierarchical Soft Mixture-of-Experts for 3D Medical Image Segmentation

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

Abstract:

In recent years, artificial intelligence has significantly advanced medical image segmentation. However, challenges remain, including efficient 3D medical image processing across diverse modalities and handling data variability. In this work, we introduce Hierarchical Soft Mixture-of-Experts (HoME), a two-level token-routing layer for efficient long-context modeling, specifically designed for 3D medical image segmentation. Built on the Mamba state-space model (SSM) backbone, HoME enhances sequential modeling through sparse, adaptive expert routing. The first stage employs a Soft Mixture-of-Experts (SMoE) layer to partition input sequences into local groups, routing tokens to specialized per-group experts for localized feature extraction. The second stage aggregates these outputs via a global SMoE layer, enabling cross-group information fusion and global context refinement. This hierarchical design, combining local expert routing with global expert refinement improves generalizability and segmentation performance, surpassing state-of-the-art results across datasets from the three most commonly used 3D medical imaging modalities and data quality.

Biography:

Szymon Płotka obtained his PhD in Computer Science in 2024 from the Informatics Institute at the University of Amsterdam, where his research focused on applying deep learning to enhance prenatal care. Szymon’s research interests lie at the intersection of computer vision, machine learning, and deep learning-based algorithms for medical image analysis. He is particularly interested in developing innovative AI-driven solutions to improve diagnostic accuracy, integrate multimodal data sources, and optimise healthcare workflows. His work aims to bridge the gap between cutting-edge artificial intelligence and real-world clinical applications, contributing to more efficient and accessible medical imaging technologies.

Piotr Wyrwiński

PUT, PCSS

Co-authors:

Krzysztof Krawiec

Poster 48: Learning to Synthesize Expressions: Symbolic Regression via Iterative Graph Expansion

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

Abstract:

Symbolic regression (SR) is a fundamental problem in machine learning and scientific discovery, where the objective is to recover an explicit analytical expression that fits observed data. Unlike conventional regression, SR does not assume a predefined model structure -- instead, it requires synthesizing both the form and parameters of a symbolic expression from scratch. In this sense, SR can be viewed as a constrained form of program synthesis from examples, where the desired output is a compact and interpretable program (in the form of a mathematical expression) that approximates an unknown target function. This task is notoriously difficult: the space of candidate programs grows exponentially with formula depth, and small syntactic changes can lead to abrupt and unpredictable changes in semantics, making the search space highly discontinuous and non-differentiable. As a result, gradient-based methods are ineffective, and even evaluating the quality of partially constructed programs can be unreliable due to their brittleness and poor generalization. These challenges mirror core issues in symbolic program synthesis, highlighting the need for learned, structure-aware guidance. To address this, we propose a neurosymbolic algorithm that frames symbolic regression as an iterative process of neural-guided graph expansion. The method incrementally expands a computational graph where each node represents a symbolic expression constructed so far, and edges reflect their compositional structure. A graph neural network (GNN) is queried throughout this process to identify which candidate subexpressions should be expanded, thereby guiding the bottom-up synthesis of increasingly complex models. To support effective guidance, each node is equipped with a learned embedding that represents the symbolic expression associated with that node. These embeddings capture both the syntactic structure of the expression (e.g., its composition and position in the graph) and its semantic behavior across training data (i.e., the outputs it produces on input examples). They are further enriched via a transformer-style attention module, which aggregates information across all input–output examples. This enables the model to form contextualized, task-aware representations of candidate expressions. The attention-guided message passing mechanism within the GNN allows the system to prioritize substructures that are not only structurally plausible but also exhibit behavior consistent with the target values. We evaluate the proposed method on a large synthetic benchmark of symbolic regression problems as well as the AI Feynman suite -- a collection of equations derived from physics. The results show that neural guidance dramatically improves success rates over unguided variants, demonstrating the benefits of solving symbolic regression iteratively and compositionally, rather than in a single step as in many purely neural approaches.

Biography:

Joanna Kaleta

SANO Poland; Warsaw University of Technology

Co-authors:

Weronika Smolak-Dyżewska, Dawid Malarz, Diego Dall'Alba, Przemyslaw Korzeniowski, Przemysław Spurek

Poster 49: PR-ENDO: Physically Based Relightable Gaussian Splatting for Endoscopy

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

Endoluminal endoscopic procedures are essential for diagnosing colorectal cancer and other severe conditions in the digestive tract, urogenital system, and airways. 3D reconstruction and novel-view synthesis from endoscopic images are promising tools for enhancing diagnosis. Moreover, integrating physiological deformations and interaction with the endoscope enables the development of simulation tools from real video data. However, constrained camera trajectories and view-dependent lighting create artifacts, leading to inaccurate or overfitted reconstructions. We present PR-ENDO, a novel 3D reconstruction framework leveraging the unique property of endoscopic imaging, where a single light source is closely aligned with the camera. Our method separates light effects from tissue properties. PR-ENDO enhances 3D Gaussian Splatting with a physically based relightable model. We boost the traditional light transport formulation with a specialized MLP capturing complex light-related effects while ensuring reduced artifacts and better generalization across novel views. PR-ENDO achieves superior reconstruction quality compared to baseline methods on both public and in-house datasets. Unlike existing approaches, PR-ENDO enables tissue modifications while preserving a physically accurate response to light, making it closer to real-world clinical use.

Biography:

Joanna Kaleta is a Computer Science graduate from Warsaw University of Technology who is currently a Ph.D. student at Sano - Center for Computational Medicine. Her research interest lies in the exploration of innovative Computer Vision applications for computed assisted surgery and diagnosis.

Michał Sala

Uniwersytet Warszawski

Co-authors:

Karol Kuźniak

Poster 50: PAST LIME: Explaining Code Generation via AST-Based Perturbations

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

Code-generation models such as Stable-Code-3B have shown strong performance in tasks like code completion, bug fixing, and refactoring. However, their black-box nature limits users’ ability to understand the factors that influence the model’s output. This lack of interpretability poses challenges in software engineering, where trustworthy and semantically coherent explanations are essential. We propose PAST LIME (Perturbed Abstract Syntax Tree LIME), a model-agnostic interpretability method that identifies which code fragments affect the probability assigned to a fixed continuation. Most existing approaches either operate at the token level without considering code structure, or attempt to align token probabilities with semantic representations without using perturbation-based analysis. In contrast, PAST LIME assigns importance scores to code fragments -- corresponding to a subset of nodes in the abstract syntax tree -- indicating whether the presence of each fragment increases or decreases the likelihood of the given continuation. PAST LIME perturbs semantically meaningful, non-overlapping code segments and queries the model’s conditional probability distribution, without requiring access to model gradients or internal parameters. Its output is independent of the sampling strategy used during generation and reflects the sensitivity of the continuation probability to structural changes in the code. Despite operating locally, the method scales to large codebases and remains computationally efficient. Beyond standard explainability, PAST LIME enables analysis of arbitrary continuations, offering new opportunities for uncovering hidden biases, evaluating model behavior, and supporting model debugging. Through experiments on Stable-Code-3B, we show that PAST LIME yields explanations more aligned with code semantics than token-level baselines, providing developers with interpretable insights into the model’s decision-making process.

Biography:

A human being. Interested in ML, distributed systems, programming languages & formal verification.

Pawel Struski

University of Warsaw

Poster 51: Competitive Market Behaviour of LLMs

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

This study explores the competitive market behaviour of Large Language Models (LLMs). We find that LLM agents exhibit systematic biases in economic decision-making tasks. In a simulated competitive market, we find that LLM buyer agents frequently bid their maximum reservation price, despite being instructed to buy at the lowest possible price. This leads to transaction prices above the competitive market equilibrium – a striking divergence from human behaviour in the same setting. We arrive at this result by replicating Smith’s (1962) classic behavioural economics experiment using LLM agents (GPT-4.1-mini-2025-04-14). We simulate a market for a fictitious good, populated by two types of agents: buyers and sellers, each with a predefined reservation price. Buyers are tasked with purchasing at the lowest possible price (but not above their reservation price), and sellers with selling at the highest possible price (but not below their reservation price). Reservation prices are structured so as to create a downward-sloping demand curve and an upward-sloping supply curve. Economic theory, supported by Smith’s original experiments, predicts that such a market should converge to a competitive equilibrium where supply equals demand. In our experiments, however, transaction prices and traded quantities consistently converge to levels above the competitive market equilibrium. Our preliminary analysis indicates that this deviation is primarily driven by non-competitive behaviour among buyer agents. Despite being instructed to maximize profits, buyers frequently bid their maximum price when asked to submit a bid, thus violating the standard assumption that agents seek to maximise utility (profits). Interestingly, this behaviour is not observed among sellers, despite near-identical (role-adjusted) prompts. This asymmetry suggests that the LLM encodes a role-based behavioural bias: associating sellers with competitive behaviour and buyers with cooperative behaviour. When this bias is accounted for, the observed price deviation aligns with theoretical predictions from economics. To our knowledge, this is the first study to apply economic theory as a tool for analysing LLM behaviour. Economic theory provides objective, consistent and tractable predictions about equilibrium outcomes in multi-agent systems. This offers a new lens for diagnosing and understanding LLM biases and decision-making patterns. Looking ahead, we plan to extend our analysis to other models (including open-source models) and apply mechanistic interpretability techniques to identify the internal representations that drive this behaviour. Ultimately, this research could inform the design and control of LLM-powered multi-agent systems, guided by principles from mechanism design – a field at the intersection of economics and game theory. References Smith, V. L. (1962). An Experimental Study of Competitive Market Behavior. Journal of Political Economy, 70(2), 111–137. https://doi.org/10.1086/258609.

Biography:

Pawel Struski is a PhD student at the University of Warsaw. His research lies at the intersection of economics and machine learning. He holds an MPhil degree in Economic Research from the University of Cambridge (2019) and a BSc degree in Economics from UCL (2018). He has previously worked as research assistant at the Institute for Fiscal Studies and as an economist in the financial sector.

Katarzyna Woźnica

Warsaw University of Technology

Co-authors:

Piotr Wilczyński, Przemysław Biecek

Poster 52: SeFNet: Linking tabular datasets with semantic feature nets

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

Tabular datasets play a significant role in wide range of machine learning applications. However, although they often address similar problems, tabular datasets are typically treated as standalone tasks. The opportunities of using previously solved problems are limited due to the lack of structured contextual information about their features and the lack of understanding of the relations between them. To overcome this limitation, we propose a new methodology called Semantic Feature Net (SeFNet), capturing the semantic meaning of the analyzed tabular features. By leveraging existing ontologies and domain knowledge, SeFNet opens up new opportunities for sharing insights between diverse predictive tasks. One such opportunity is the Dataset Ontology-based Semantic Similarity (DOSS) measure, which quantifies the similarity between datasets using relations across their features. In this paper, we present an example of SeFNet’s application prepared for a collection of predictive tasks in healthcare, with the features’ relations derived from the SNOMED-CT ontology. The proposed SeFNet methodology and the accompanying DOSS measure address the issue of limited contextual information in tabular datasets. By incorporating domain knowledge and establishing semantic relations between features, we enhance the potential for meta-learning and enable valuable insights to be shared across different predictive tasks.

Biography:

Katarzyna is a researcher at Warsaw University of Technology with a PhD in Computer Science. Her work centers on AutoML, hyperparameter optimization, and meta-learning, with a focus on human-in-the-loop methods and Automated Data Science. She collaborates with physicians to develop machine learning solutions for medical applications.

Stanisław Pawlak

Warsaw University of Technology

Co-authors:

Jan Dubiński, Daniel Marczak, Bartłomiej Twardowski

Poster 53: Backdoor Vectors: a Task Arithmetic View on Backdoor Attacks and Defenses

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

Model merging (MM) recently emerged as an effective method for combining large deep learning models. However, it poses significant security risks. Recent research shows that it is highly susceptible to backdoor attacks, which introduce a hidden trigger into a single fine-tuned model instance that allows the adversary to control the output of the final merged model at inference time. In this work, we propose a simple framework for understanding backdoor attacks by treating the attack itself as a task vector. Backdoor Vector (BV) is calculated as the difference between the weights of a fine-tuned backdoored model and fine-tuned clean model. BVs reveal new insights into attacks understanding and a more effective framework to measure their similarity and transferability. Furthermore, we propose a novel method that enhances backdoor resilience through merging dubbed Sparse Backdoor Vector (SBV) that combines multiple attacks into a single one. We identify the core vulnerability behind backdoor threats in MM: inherent triggers that exploit adversarial weaknesses in the base model. To counter this, we propose Injection BV Subtraction (IBVS) -- an assumption-free defense against backdoors in MM. Our results show that SBVs surpass prior attacks and is the first method to leverage merging to improve backdoor effectiveness. At the same time, IBVS provides a lightweight, general defense that remains effective even when the backdoor threat is entirely unknown.

Biography:

Stanisław Pawlak is a Ph.D. student and AI researcher working at Warsaw University of Technology. He received a M.Sc. degree in data science and a B.Sc. in applied computer science from the Warsaw University of Technology. Stanisław coauthored multiple publications on world top-tier AI conferences, including NeurIPS 2023 and CVPR 2024. He also worked as a programmer, AI engineer building ML-powered applications, and AI consultant. His research focuses on continual learning, generative models, and ML security. His latest efforts aim to understand backdoors in model merging learning paradigm.

Bartosz Jezierski

Warsaw University of Technology

Co-authors:

Mateusz Jarosz, Vladimir Zaigrajew

Poster 54: Comparing Different Jailbreak Detection Methods

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

As Large Language Models (LLMs) are increasingly deployed in production, their vulnerability to novel 'jailbreak' attacks poses a significant security risk. Popular safeguards often fail to generalize to these unseen threats, creating a critical security gap. We evaluate this generalization failure by comparing prominent open-source safeguards against targeted, data-efficient fine-tuning of an encoder only (ModernBERT) and a decoder only (Gemma 3) model. Our experiments on the established benchmark indeed confirm the critical generalization gap against unseen jailbreaks. Guardrails like Llama Guard 2 and 3 achieve less than 55% accuracy on the unseen RedTeams2K dataset. We discover that resource-efficient fine-tuning models yield superior performance compared to baseline open-source safeguards. While even simple head-tuning of ModernBERT surpasses the baseline, more comprehensive LoRA fine-tuning on Gemma yields the best results. However, the largest performance gain, which boosts detection accuracy to over 95%, comes from introducing a small, curated set of just 200 similar jailbreak examples. By breaking down performance by category, our analysis exposed a significant, shared blind spot. We found that detecting prompts designed to violate privacy was a critical challenge that affected every model we tested. Our findings demonstrate that effective LLM safety is found not in large, static guardrails, but in an agile strategy that prioritizes rapidly fine-tuning smaller, general-purpose models on new, similar attack patterns as they appear.

Biography:

Data Science student at Warsaw University of Technology.

Paulina Kaczyńska

IPPT PAN

Co-authors:

Nazanin Amirinejad, Tomasz Lipniacki

Poster 55: Immunogenic cell death detection and classification with Machine Learning

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

Viral infection can lead to regulated cell deaths that may have both non-immunogenic (apoptosis) or immunogenic character (pyroptosis or necroptosis). As immunogenic deaths allow adaptive immune responses to develop, differentiating and describing these processes would give us important insights to the development of immunogenic responses during various viral infections. Current tools predominantly detect apoptosis or general cell death, with relatively few methods capable of distinguishing between specific types. In particular, pyroptosis remains underexplored. Some recent approaches attempt to differentiate between apoptosis and ferroptosis or necrosis, but these typically operate at the image level rather than at single-cell resolution. Moreover, most existing methods rely on fluorescent markers that either translocate to dying cells or respond to caspase activity specific to a given death pathway. This fluorescence-based approach is constrained by the practical limit of using only a few simultaneous markers, reducing the ability to visualize other critical aspects of virus-cell interactions. To address these limitations, we developed a dataset for detecting and classifying apoptosis and pyroptosis based on cell morphology in brightfield microscopy images. Pyroptotic cells were manually annotated by experts using morphological criteria, while propidium iodide staining provided reference information on cell death. The dataset supports two use cases: image classification (single cells and their immediate surroundings) and object detection with bounding boxes. All data were collected from live-cell time-lapse microscopy, enabling temporal analysis. Quantitative features extracted from the cells are also available in tabular format. We evaluate the performance of deep convolutional neural networks on this dataset—first in the classification setting and then in the object detection task. Finally, we explore several strategies for incorporating temporal information to improve detection and classification performance over time.

Biography:

Paulina is currently in the first year of her PhD at the Institute of Fundamental Technological Research of the Polish Academy of Sciences (IPPT PAN). Her research focuses on modeling regulatory networks at single-cell resolution using machine learning methods. She holds a Bachelor's degree in Physics and a Bachelor's degree in Cognitive Science, both obtained through the College of Inter-Faculty Individual Studies in Mathematics and Natural Sciences (MISMaP) at the University of Warsaw. She later completed a Master's degree in Machine Learning at the University of Warsaw, where her thesis explored the visualization of node features using Accumulated Local Effects in Graph Neural Networks.

Ryszard Staruch

Adam Mickiewicz University

Co-authors:

Filip Graliński, Daniel Dzienisiewicz

Poster 56: Adapting LLMs for Minimal-edit Grammatical Error Correction

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

Decoder-only large language models have shown superior performance in the fluency-edit English Grammatical Error Correction, but their adaptation for minimal-edit English GEC is still underexplored. To improve their effectiveness in the minimal-edit approach, we explore the error rate adaptation topic and propose a novel training schedule method. Our experiments set a new state-of-the-art result for a single-model system on the BEA-test set. We also detokenize the most common English GEC datasets to match the natural way of writing text. During the process, we find that there are errors in them. Our experiments analyze whether training on detokenized datasets impacts the results and measure the impact of the usage of the datasets with corrected erroneous examples. To facilitate reproducibility, we have released the source code used to train our models.

Biography:

Ryszard is a PhD student at Adam Mickiewicz University and a machine learning researcher at the Center for Artificial Intelligence AMU. His main research areas are grammatical error correction and information retrieval. He is the author of papers presented at international NLP conferences and the winner of the multilingual grammatical error correction shared task MultiGEC-2025.

Łukasz Niedźwiedzki

University of Warsaw/Taxus IT

Co-authors:

Abigail Turner, Mononito Goswami, Artur Dubrawski

Poster 57: Exploiting Spatial Structure with Time Series Foundation Models

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

Abstract:

Multivariate time series data is pervasive in real-world applications, from climate monitoring and traffic forecasting to clinical diagnostics. Despite this, current foundation models like MOMENT often fall short in capturing the complex structure inherent to such data. In particular, MOMENT processes each channel independently during operations like patching, which limits its ability to model the rich inter-variable relationships that are often critical in real-world settings. Many domains exhibit strong spatial or functional dependencies between variables - for example, between nearby sensors in a physical environment - which are ignored by treating channels in isolation. This project addresses these limitations by integrating graph neural networks with MOMENT to explicitly model relationships between channels. By embedding spatial and inter-variable dependencies into the model architecture, we aim to better capture the underlying structure of multivariate time series data. This integration introduces structural inductive biases that encourage the model to learn more meaningful representations and to generalize more effectively across different datasets and domains. We conduct extensive evaluations to assess both the predictive performance and the internal representations learned by the model. Through a variety of analyses, we investigate how incorporating graph-based relational modeling impacts the way MOMENT encodes and utilizes cross-variable information, shedding light on the mechanisms that drive improved performance.

Biography:

Łukasz Niedźwiedzki is a Master's student at the Faculty of Physics, University of Warsaw. He also works as a Machine Learning Engineer at Taxus IT and actively participates in research at the Auton Lab at Carnegie Mellon University and the MI2 Lab at Warsaw University of Technology.

Hubert Plisiecki

Polish Academy of Science

Co-authors:

Paweł Lenartowicz, Artur Pokropek, Maria Flakus

Poster 58: Words Apart: Mapping Psychological Differences Through Semantic Space

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

This study introduces a new method for analyzing how word meanings shift across psychological groups, using 1300 essays from participants who also completed measures of attitudes like collective narcissism and trust in science. We examine how people high versus low on these psychological dimensions use language differently when discussing topics such as national identity, migration, and climate change. Our word2vec-based technique identifies how key concepts take on different meanings between contrasting groups by comparing word associations in their respective language samples. By retro-fitting target words based on an external pretrained word2vec model, we bypass the common word2vec data sparsity problem, allowing us to capture meaningful shifts in language use. The method provides two valuable insights for psychological research: (1) statistical inference (p-values) of semantic differences between groups through, and (2) visualization of how concepts relate differently across psychological divides by examining nearby words in each group's semantic space. This method expands the applications of machine learning based NLP methods within the realm of social sciences, offering a new window into how individual differences manifest in language use, but can also serve as inspiration for repurposing older NLP methods to new tasks.

Biography:

Hubert Plisiecki is an AI researcher specializing in the application of machine learning and NLP to research in social sciences, particularly in modeling psychological phenomena in text. Currently completing his PhD at the Polish Academy of Sciences, Hubert has authored multiple publications on innovative machine learning approaches in social sciences, including articles in journals such as Behavior Research Methods. He is also a co-founder of the Society for Open Science, promoting research quality and transparency in Poland and abroad.

Dawid Plaskowski

Auton Lab @ Carnegie Mellon Univeristy / Łukasiewicz Research Network

Co-authors:

Willa Potosnak, Michał Williński, Artur Dubrawski

Poster 59: Beyond Context Limits: Retrieval-Augmented Time Series Forecasting with Prior-data Fitted Networks and Time Series Foundation Models

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

Abstract:

In-context learning has emerged as a powerful paradigm in large language models, enabling few-shot generalization from carefully chosen examples. We are interested in investigating this capability beyond language, focusing on the generalization of Prior-data Fitted Networks (PFNs), pretrained only on synthetic data for tabular tasks, to time series. Prior work on tabular foundation models such as TabPFN, applied to time series forecasting tasks, shows that on GIFT-eval only ~40% of the context window is effectively used, and adding more in-context examples can sharply degrade performance. This motivates the question: can retrieval better utilize context, matching or exceeding performance with fewer examples and extending the effective context window through targeted selection? We propose a retrieval-augmented forecasting framework for tabular foundation models such as TabPFN and Mitra adapting principles from Retrieval-Augmented Generation (RAG) in LLMs. Our approach uses MOMENT, a time-series foundation model, as the retrieval module, drawing from both domain-specific sources and large-scale corpora such as TimeSeriesPile to select semantically relevant time series at test time, which are then incorporated into the model’s in-context examples.

Biography:

I am an aspiring researcher exploring the inner workings of neural networks through interpretability and neural scaling laws. My experience spans applied machine learning in both academic and industrial settings, including two research internships at Carnegie Mellon University’s Robotics Institute Summer Scholars program, work at the Łukasiewicz Research Network on natural language processing projects, and contributions at a drone technology startup.

Julian Kędys

Poznan Supercomputing and Networking Center, PAS

Co-authors:

Cezary Mazurek

Poster 60: Interpretable Coupling Structure Beyond Deep Learning: Probabilistic and Energy-Based Modelling of Multivariate (Neural) Dynamics

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

poster

Abstract:

Many scientific problems present high-dimensional, noisy and small-N observations where deep nets can be data-hungry and opaque. I will present a set of methods that combines energy-based models and probabilistic/Bayesian inference to model multivariate dynamics with interpretable coupling structure. Concretely, we use the pairwise maximum-entropy model (PMEM) - the Ising/Boltzmann machine without hidden units - as an undirected Markov random field over binary (or binarised) variables, giving a principled, minimum-assumption distribution that matches empirical first- and second-order statistics. Methodologically, I will cover three complementary estimators that trade off bias, variance and compute: (i) exact maximum likelihood via full state enumeration for small systems; (ii) a scalable pseudo-likelihood optimiser with safeguarded line search and L2 regularisation; and (iii) a variational Bayes formulation that places Gaussian priors on fields and couplings to yield posterior means/precisions and credible intervals-useful when data are limited. Sampling-based MCMC (Numba-accelerated Metropolis, adaptive multi-chain with R-hat) serves both for model checking and for deriving kinetic summaries. Beyond fitting, the methodology turns parameters into mechanistic diagnostics. We compute phase-diagram analyses in (μ,σ) space of the coupling matrix to place subjects/groups relative to critical regimes (qualitative characterisation of the system's dynamics' organisation); derive energy landscapes and disconnectivity graphs that expose stable states and transition barriers; and construct state-to-state transition matrices, dwell times and transition pathways using single-spin-flip kinetics, providing a kinetic extension of the equilibrium model. Bootstrap + ridge shrinkage give robust control baselines, while balanced objective weights stabilise subject positioning. Strengths include (a) interpretability by design (explicit couplings and energy), (b) small-data suitability (max-entropy constraints, VB uncertainty, PL scalability), and (c) generative capability (the model defines p(s); MCMC draws synthetic trajectories). I will also outline practical limitations-e.g., binarisation assumptions, scaling to very large N, and equilibrium vs. real-world dynamics-and discuss mitigations (regularisation, block bootstrap uncertainty, kinetic extensions, and when deep learning is preferable). Although the talk will be illustrated on resting-state neural dynamics in neurodevelopmental cohorts - mapping altered basins, abnormal couplings between regions, and temporal signatures - the methodology is domain-agnostic. It applies to any system with meaningful binary (or discretised) states: genomics (gene on/off), proteomics contacts, ecology, and complex systems in physics or social science. The focus is on the general principles and reusable methods, not domain specifics. The session will include a short illustration of the visual analytics (phase diagrams; probability fit indices; barrier distributions) and several engineering improvements (adaptive samplers; robust contour-based positioning) that make these classical models practical for modern hypothesis-testing and ML-enabled research workflows. Overall, the talk advocates a probabilistic and energy-based alternative - complementary to deep learning - whenever interpretability, uncertainty, and mechanistic structure are primary goals, including in scientific contexts.

Biography:

Julian earned a B.Sc. in Artificial Intelligence from Poznań University of Technology in 2024, where he was an active member of the GHOST (“Group of Horribly Optimistic STatisticians”) Student Scientific Group. During the programme, he spent two semesters at the University of Luxembourg, where he also completed a three-month research placement in applied reinforcement learning and computer vision. Since graduating, Julian has been building on his machine-learning background by pivoting into computational neuroscience. In late 2024, he undertook a three-month research internship at the International Research Center for Neurointelligence (IRCN), University of Tokyo, studying abnormal brain dynamics in neurodevelopmental conditions, including autism spectrum disorder (ASD), through latent-representation analysis and mathematical modelling with tools drawing on statistical physics and machine learning. Since December 2024, he has been affiliated with the Poznań Supercomputing and Networking Center (PSNC), PAS, Department for Digital Medicine. His work spans computational neuroscience, deep-learning applications for computational biology and medicine, and exploratory analyses of quantum-computing applications in the life sciences and engineering.

Kajetan Dymkiewicz

University of Cambridge

Co-authors:

Ivan Vulic, Helen Yannakoudakis, Eilam Shapira, Roi Reichart

Poster 61: Unpacking the Potential: Refining and Evaluating Language Models Across Dimensions

Friday / 17 October 12:15 - 13:45 (Poster Session 2)

Abstract:

Large Language Models (LLMs) demonstrate strong performance across tasks and languages, but the mechanisms of cross-dimensional transfer, how gains in one capability, task, or language influence others, remain poorly understood. In this work, we consider four such dimensions - language, task type, model family, and model size - and systematically evaluate transfer patterns across them. We present a systematic evaluation of transfer patterns across seven state-of-the-art models from the LLaMA 3 and Qwen 2.5 families, spanning sizes from 0.5B to 8B parameters. Using multilingual benchmarks covering reasoning, factuality, coding, fairness, and math, we fine-tune models on individual task–language pairs and evaluate their performance across the entire evaluation space. Our analysis reveals a consistent asymmetry: within-task, cross-lingual transfer is broadly beneficial, while cross-task transfer frequently incurs collateral harm. Transfer is structured by donor–recipient roles: certain hub languages (e.g., Turkish, Dutch) and tasks (e.g., Factuality, Fairness) export substantial gains but often degrade others, whereas Coding and Math act as comparatively benign donors yet brittle recipients. The results suggest that effective fine-tuning strategies must explicitly account for these dynamics, balancing on-task improvements with the risk of off-task degradation to unlock the full potential of LLMs across languages and capabilities.

Biography:

Kajetan is a PhD student at the Language Technology Lab at University of Cambridge. He obtained his MSc in AI at King's College London and a BSc in Computer Science at Wrocław University of Science and Technology. His research focuses on the intersection of LLM efficiency and AI Safety.

/ Student Research Workshop (SRW) Talks

Aleksandra Gwiazda

Warsaw University of Technology

Co-authors:

Paulina Szymczak, Ewa Szczurek

SRW Talk 1: PepGlider: Property Regularized VAE for interpretable and controllable peptide design

Wednesday / 15 October 8:10 - 8:30 Main Hall (Student Research Workshop)

slides slides (online preview)

Abstract:

Antibiotic resistance is a growing global health threat, driving the need for alternative therapeutics. Antimicrobial peptides (AMPs) offer a promising solution due to their broad-spectrum activity and lower propensity to induce resistance. While deep learning–based generative models, such as variational autoencoders (VAEs), have been explored for designing novel AMPs, their limited controllability and interpretability hinder practical application. In this work we present Attribute-Regularized VAE (AR-VAE), a generative model that explicitly incorporates key peptide attributes directly into the training objective. A regularization loss is applied to structure the latent space, enabling controlled traversal along interpretable directions corresponding to charge, length, and hydrophobicity. Experimental results demonstrate that AR-VAE achieves improved latent space disentanglement and enhanced performance on metrics of interpretability, modularity, and predictability. The model effectively captures and manipulates properties correlated with antimicrobial activity, enabling targeted AMP generation. AR-VAE offers a principled and interpretable approach for AMP design, addressing the urgent problem of AMR through targeted AMP generation.

Biography:

Aleksandra Gwiazda obtained her Bachelor's degree in Biomedical Engineering from the Warsaw University of Technology in February 2023. She is currently pursuing a specialization in Artificial Intelligence as part of the Informatics program at WuT. Aleksandra is collaborating with Prof. Ewa Szczurek's research group at Helmholt.

Łukasz Sztukiewicz

Independent Researcher

Co-authors:

Ignacy Stępka, Michał Wiliński, Jerzy Stefanowski

SRW Talk 2: DetoxAI - Python Package for Debiasing Neural Networks

Wednesday / 15 October 8:30 - 8:50 Main Hall (Student Research Workshop)

Abstract:

Despite growing awareness of fairness in machine learning, the lack of dedicated debiasing techniques and practical software frameworks remains a major barrier - especially in vision tasks. To address this, we introduce DetoxAI, a Python-based framework for post-hoc debiasing of neural networks in image classification. Built with deep learning workflows in mind, DetoxAI combines interventions, fairness metrics, and visualization tools into a single, production-ready package. Our method uses post-training adaptations to reduce bias without degrading model accuracy. By operating on high-level semantic features, DetoxAI tackles the unique challenge of vision models, where sensitive attributes like race or gender are rarely explicitly represented. The toolkit offers a modular and accessible interface, making it suitable for real-world use across different domains. Through experiments, we show that DetoxAI consistently improves the fairness-accuracy trade off over standard models. Attribution map analyses further confirm that DetoxAI reduces reliance on protected features. Parts of this work were published and presented at XAI Late-Breaking Work 2025, KDD UMC 2025, and the ECML Demo Track 2025.

Biography:

Łukasz Sztukiewicz holds a Bachelor of Science degree in Artificial Intelligence from Poznan University of Technology. He was a research fellow at Carnegie Mellon University and worked as a machine learning engineer at Molecule.one.

Mikołaj Janusz

Jagiellonian University

Co-authors:

Adam Wróbel, Bartosz Zieliński, Dawid Rymarczyk

SRW Talk 3: OMENN: One Matrix to Explain Neural Networks

Wednesday / 15 October 8:50 - 9:10 Main Hall (Student Research Workshop)

slides slides (online preview)

Abstract:

Deep Learning (DL) models are often black boxes, which makes their decision-making processes difficult to interpret. This lack of transparency has driven advancements in eXplainable Artificial Intelligence (XAI), a field dedicated to clarifying the reasoning behind DL model predictions. Among these, attribution-based methods such as LRP and GradCAM are widely used, though they rely on approximations that can be imprecise. To address these limitations, we introduce One Matrix to Explain Neural Networks (OMENN), a novel post-hoc method that represents a neural network as a single, interpretable matrix for each specific input. This matrix is constructed through a series of linear transformations that represent the processing of the input by each successive layer in the neural network. As a result, OMENN provides locally precise, attribution-based explanations of the input across various modern models, including ViTs and CNNs. We present a theoretical analysis of OMENN based on dynamic linearity property and validate its effectiveness with extensive tests on two XAI benchmarks, demonstrating that OMENN is competitive with state-of-the-art methods. The work is available on Arxiv.

Biography:

Mikołaj Janusz is a Computer Science M.Sc. student at the Jagiellonian University and a Student Researcher at GMUM (gmum.net). During his studies, he conducted research on XAI for drug discovery as a student participant in the FIRST TEAM FENG program (grant "Interpretable and Interactive Multimodal Retrieval in Drug Discovery"), various topics in computer vision, and pruning regimes. Beyond his research, he gained practical software engineering experience in distributed systems and compute infrastructure during internships at Google, Meta, Amazon, and the systematic-trading firm Quadrature. Currently, he's mainly interested in topics of compute, applied machine learning, and software engineering.

Tomasz Dądela

Jagiellonian University

Co-authors:

Adam Kania, Przemysław Spurek, Maciej Rut

SRW Talk 4: From Blurry to Sharp: Enabling High-Frequency Detail in Implicit Neural Representations

Wednesday / 15 October 9:10 - 9:30 Main Hall (Student Research Workshop)

Abstract:

Implicit Neural Representations have recently gained attention as a powerful approach for continuously representing signals – such as images, videos, and 3D shapes – using multilayer perceptrons (MLPs). However, MLPs are known to exhibit a low-frequency bias, which limits their ability to capture high-frequency details accurately. Our aim is to find optimal parameters using a compute-efficient method, without the need to perform multiple rounds of representation learning. Such a configuration should enable the model to learn the details of the representation more quickly. We explore the relationship between model performance and the frequency of the embedding, the frequency of the target image, as well as key MLP design choices such as the number of features and layers. Alongside experiments with real-world images, we created a synthetic dataset to identify individual factors that determine the performance of a particular model.

Biography:

Tomasz Dądela is a final year Machine Learning M.Sc. student at the Jagiellonian University. Currently, he is actively involved in projects with the Group of Machine Learning Research (GMUM) at the Jagiellonian University.

Rafał Malcewicz

Carnegie Mellon University, GHOST Day

Co-authors:

Ignacy Stępka, Abby Turner, Artur Dubrawski

SRW Talk 5: Semantic Label Reconstruction: How to breach privacy in Federated Learning

Wednesday / 15 October 9:50 - 10:10 Main Hall (Student Research Workshop)

slides

Abstract:

Federated Learning enables clients to collaboratively train models by sharing gradients with a central server or with each other, offering both computational efficiency and enhanced privacy compared to sharing raw data. However, prior research has demonstrated that gradient inversion attacks can partially reconstruct the original input data from shared gradients. Unfortunately, this approach faces significant limitations, particularly in high-batch-size settings, due to the inherent information loss during gradient computation, which severely degrades reconstruction quality. In this work, we propose a shift in focus from input reconstruction to task reconstruction, specifically targeting the recovery of semantic labels from shared gradients. While this form of attack does not reveal full input data, it still constitutes a meaningful privacy breach. Importantly, semantic label reconstruction remains more robust under increasing batch sizes compared to full data reconstruction, making it a viable and concerning threat vector in practical federated learning scenarios.

Biography:

Rafał Malcewicz is pursuing a Bachelor of Science and Technology at Aalto University. He has research experience at the Auton Lab, Carnegie Mellon University, where he worked on gradient inversion attacks in the federated learning setting. Currently, he serves as the Project Leader of GHOST Day AMLC (Applied Machine Learning Conference).

Nikodem Świerkowski

Wrocław University of Science and Technology

Co-authors:

Daniel Borkowski, Mikołaj Jastrzębski

SRW Talk 6: Filling in the Blanks: Redefining Inpainting with a Novel Geometrical Latent Space VAE Architecture

Wednesday / 15 October 10:10 - 10:30 Main Hall (Student Research Workshop)

slides slides (online preview)

Abstract:

Image inpainting — reconstructing missing parts of an image — is a crucial task in computer vision with applications in photo restoration, image editing, and beyond. This task requires generative models that not only produce visually coherent outputs but also reason about the uncertainty and semantic context inherent in incomplete data. While lightweight convolutional autoencoders and GANs are commonly used due to their simplicity and training speed, they lack a probabilistic treatment of uncertainty and limit the interpretability and exploration of latent space. In contrast, Variational Autoencoders (VAEs) offer a more expressive framework for modeling the data distribution and generating diverse, plausible reconstructions. We compare four VAE-based models: the standard VAE, VQ-VAE (which uses latent space quantization), VAE-GAN (incorporating an adversarial component), and TreeVI — a model proposed at NeurIPS 2024 that captures inter-sample correlations using a tree-structured latent space built via a Minimum Spanning Tree (MST). While TreeVI improves reconstruction quality, its computational overhead and non-differentiable graph structure make it difficult to scale in practice. To address these limitations, we introduce GeoVAE — a novel model that treats the latent space as a graph, where edges represent correlations between dimensions. Unlike TreeVI, GeoVAE learns these correlations dynamically through two graph convolutional networks (GNNs): GNNinst, which captures inter-sample relationships, and GNNdim, which estimates the importance of individual latent dimensions. The reparameterization trick is extended into an adaptive mechanism that allows for scalable, fully differentiable correlation modeling. This enables learning on true correlation structures, rather than approximating them through trees, showcasing the potential of integrating geometric learning into standard deep generative models. We evaluate all models on a dense inpainting task involving 64×64 RGB images with randomly placed square masks of varying size and location. The models are trained to reconstruct the missing regions conditioned on the visible context. GeoVAE achieves the lowest MSE across all models, with competitive SSIM and PSNR. Notably, its per-epoch training time is only ~1.5× that of the base VAE and significantly faster than TreeVI, while delivering better semantic and perceptual coherence.

Biography:

Nikodem Świerkowski is a Master's student in Artificial Intelligence at Wrocław University of Science and Technology. He holds a Bachelor's degree in Applied Computer Science from the same university. His current research focuses on enhancing the effectiveness of state-of-the-art video plagiarism detection methods. Enthusiast of probabilistic graphs models, explainable AI and cooking :)

Paulina Hładki

Poznan University of Technology

Co-authors:

Marek Justyna, Maciej Antczak, Marta Szachniuk

SRW Talk 7: GraphaFold: Graph Neural Network for predicting non-canonical RNA base pairing

Wednesday / 15 October 10:30 - 10:50 Main Hall (Student Research Workshop)

slides slides (online preview)

Abstract:

Recent breakthroughs in biomedicine, such as mRNA vaccines and gene-editing technologies, have highlighted the critical role of RNA molecules in health and disease. RNA (ribonucleic acid) is not just a carrier of genetic information; it also regulates cellular processes, assembles complex molecular machines, and can even act as a biological catalyst. The function of RNA depends largely on its three-dimensional structure, which is formed through networks of hydrogen bonds - similar to magnets holding different parts of the molecule together. These bonds create base pairs, which come in two types: canonical (making up about 70% of all base pairs in RNA) and non-canonical (alternative, which can constitute up to 30%). While most computational tools focus on canonical base pairs, non-canonical interactions are crucial for shaping RNA's structure, especially in flexible regions and in contacts that determine how RNA interacts with proteins and other molecules. However, predicting these non-canonical base pairs from sequence data remains a major challenge, as they are highly variable and less well-represented in existing training data. In this work, we introduce GraphaFold, a graph neural network (GNN)-based approach designed to predict non-canonical RNA base pairs. Our method starts with known canonical interactions - either from experiments or existing software - and represents the RNA molecule as a graph, where each nucleotide is a node and possible interactions are edges. We use advanced neural embeddings, generated by a pretrained RNA language model, to provide rich input features. The GNN learns to recognize complex patterns associated with non-canonical base pairing by leveraging features of the sequence, structural context, and graph connectivity. To address the problem of class imbalance (since non-canonical pairs are relatively rare), we use a reweighted loss function during training. We created a comprehensive dataset of RNA fragments from high-resolution 3D structures, capturing a wide variety of non-canonical motifs, to train and rigorously test the model. Benchmarks show that GraphaFold outperforms current deep learning tools in predicting non-canonical interactions, while scaling well to different RNA types and sizes. The method offers a practical solution for advancing RNA structure modeling, as GraphaFold’s predictions can enrich 2D structural maps and serve as valuable constraints for more detailed 3D modeling. Funding: This research was supported by grant 2024/53/B/ST6/02789 from the National Science Centre, Poland.

Biography:

Paulina Hładki is a master’s student in computer science with a specialization in artificial intelligence at Poznan University of Technology. She is currently working on deep learning models for RNA structure prediction. With experience in both academic research and international IT environments, she bridges theoretical insight with practical machine learning applications.

Łukasz Janisiów

Jagiellonian University

Co-authors:

Marek Kochańczyk, Bartosz Zieliński, Tomasz Danel

SRW Talk 8: Enhancing Chemical Explainability Through Counterfactual Masking

Wednesday / 15 October 10:50 - 11:10 Main Hall (Student Research Workshop)

slides slides (online preview)

Abstract:

Molecular property prediction is a crucial task that guides the design of new compounds, including drugs and materials. While explainable artificial intelligence methods aim to scrutinize model predictions by identifying influential molecular substructures, many existing approaches rely on masking strategies that remove either atoms or atom-level features to assess importance via fidelity metrics. These methods, however, often fail to adhere to the underlying molecular distribution and thus yield unintuitive explanations. In this work, we propose counterfactual masking, a novel framework that replaces masked substructures with chemically reasonable fragments sampled from generative models trained to complete molecular graphs. Rather than evaluating masked predictions against implausible zeroed-out baselines, we assess them relative to counterfactual molecules drawn from the data distribution. Our method offers two key benefits: (1) molecular realism underpinning robust and distribution-consistent explanations, and (2) meaningful counterfactuals that directly indicate how structural modifications may affect predicted properties. We demonstrate that counterfactual masking is well-suited for benchmarking model explainers and yields more actionable insights across multiple datasets and property prediction tasks. Our approach bridges the gap between explainability and molecular design, offering a principled and generative path toward explainable machine learning in chemistry.

Biography:

Łukasz Janisiów is a first-year doctoral student in the Group of Machine Learning Research at Jagiellonian University and a participant in the ELLIS PhD Program. His research focuses on the application of explainable artificial intelligence (XAI) in drug discovery.

/ Tutorials

Oleksii Furman

Wrocław University of Science and Technology

Łukasz Lenkiewicz

Wrocław University of Science and Technology

Marcel Musiałek

Wrocław University of Science and Technology

Tutorial 1: Counterfactual Explanations: From Theory to Implementation

Saturday / 18 October 9:00 - 13:00

Description:

This tutorial demystifies Counterfactual Explanations, a powerful technique for making black-box AI models transparent and actionable. Unlike traditional XAI methods that simply highlight important features, counterfactuals provide practical, user-friendly answers to the essential question: “What needs to change to get a different outcome?” In this tutorial, participants will progress from theoretical foundations to implementing counterfactual methods and using them. You’ll learn how to generate explanations that help end-users understand not just why they were denied a loan, but precisely what they need to change to get approved next time. The tutorial balances theory with extensive coding exercises. We’ll cover both basic implementations and advanced techniques using modern methods. By the end, you’ll understand how to use counterfactual explanations for your AI systems to improve transparency. You’ll leave with ready-to-use code and practical implementation strategies.

Knowledge Prerequisites

Intermediate Python programming skills (comfortable with functions, classes, and common data structures)
Basic understanding of machine learning concepts (classification, features, training/testing)
Familiarity with common libraries like NumPy, Pandas, and Matplotlib
Basic knowledge of classification models (particularly decision trees and random forests)
Exposure to the concept of model explainability (though expertise is not required)

Technical Setup: Participants should have the following installed prior to the tutorial: Python 3.10+ environment (Anaconda distribution recommended) and Jupyter Notebook or JupyterLab.

Biography:

Oleksii Furman is a Ph.D. student in Artificial Intelligence at Wrocław University of Science and Technology, specializing in generative models and explainable AI (XAI). His research develops innovative approaches to enhance the transparency and interpretability of black-box machine learning models, particularly through counterfactual explanations. He aims to bridge the gap between complex AI systems and human understanding, making machine learning more accessible and trustworthy. He has co-authored papers at prominent machine learning conferences.

Łukasz Lenkiewicz is pursuing his Ph.D. in Artificial Intelligence at Wrocław University of Science and Technology, where he investigates new directions in explainable AI. His doctoral research focuses on developing counterfactual explanations at the local, global, and group-wise level, aiming to capture both individual decisions and broader model behavior. Beyond tabular data, he explores explainability in computer vision, developing techniques that clarify how visual recognition models operate and can be put to use. His research highlights practical integration, focusing on turning theoretical advances in explainability into real-world applications.

Marcel Musiałek is a third-year undergraduate student at Wrocław University of Science and Technology, with a strong interest in computer vision, particularly in the medical sector and explainable AI (XAI). He actively develops his research interests through scientific projects and throughout the university’s science club. His focus is on expanding his knowledge and building solid foundations for future education and research.

Maura Pintor

University of Cagliari

Giorgio Piras

University of Cagliari

Tutorial 2: Where ML Security is Broken and How to Fix it

Saturday / 18 October 9:00 - 13:00

Description:

Adversarial robustness is a critical concern for modern machine learning systems, yet reliably evaluating model resilience to adversarial attacks remains a major challenge. In practice, robustness is often measured using gradient-based methods that optimize perturbations to simulate worst-case inputs. These empirical evaluations might provide an inaccurate picture of model security, as small flaws in the attack setup can lead to overly optimistic results. Without systematic testing and diagnostic tools, even well-intentioned evaluations risk repeating past mistakes.

This tutorial will guide participants through the fundamentals of adversarial robustness evaluation, emphasizing the importance of rigorous and reproducible methods. We will present a comparison framework for benchmarking gradient-based attacks that helps identify optimization failures, standardize evaluation protocols, and support fair comparisons across models and datasets. Through practical examples, attendees will learn how to design, test, and debug adversarial attacks in a principled way. Looking forward, we discuss how these challenges extend to the evaluation of foundation models and multimodal systems, where conventional adversarial techniques may be adapted and still produce significant impact. As machine learning models continue to grow in scale and complexity, there is a growing need for robust and efficient evaluation methodologies. We conclude with a discussion of open problems and future research directions aimed at strengthening the reliability and trustworthiness of machine learning under adversarial conditions.

Prerequisites: Participants should have a basic understanding of machine learning and deep learning concepts, including neural network training and evaluation. Familiarity with PyTorch is strongly recommended, as examples and hands-on exercises will be based on it. Some background in adversarial machine learning (e.g., knowledge of adversarial examples or gradient-based attacks) is helpful but not strictly required, as essential concepts will be introduced during the tutorial.

Software Requirements: All hands-on exercises will be conducted using Google Colab, so no local installation is necessary. Participants only need a Google account and a modern web browser. We will provide pre-configured Colab notebooks with all dependencies already listed, and they will be installed during the tutorial. Access to a GPU runtime in Colab is recommended for optimal performance but not strictly required.

Biography:

Maura Pintor (PhD 2022, honors) is an Assistant Professor at the PRA Lab, University of Cagliari, Italy. Her research focuses on optimizing and debugging adversarial robustness evaluations. She has held visiting positions at the University of Tuebingen (Germany, 2020), SCCH (Austria, 2021), and the Computer Vision Center (Spain, 2024). Maura serves as Area Chair for NeurIPS, Associate Editor for Pattern Recognition, and regularly reviews for top-tier conferences and journals, including ACM CCS, ICLR, ECCV, and ICCV. She is co-chair of the ACM Workshop on Artificial Intelligence and Security (AISec), co-located with ACM CCS, and contributes to several EU Horizon projects, including ELSA, Sec4AI4Sec, and CoEvolution. She is the main maintainer of the open-source SecML-Torch library.

Giorgio Piras (PhD 2025, honors) is a Postdoctoral Researcher at the PRA Lab, University of Cagliari, Italy. His research interests broadly cover adversarial machine learning, with a particular focus on adversarial pruning methods and LLM security. During his PhD, he was a visiting student at the Karlsruhe Institute of Technology, Germany. He regularly serves as a reviewer for Pattern Recognition, Neurocomputing, and IEEE TIFS journals, and AAAI, USENIX, ACM CCS conferences. He is now involved with the University of Cagliari in the EU Horizon Projects Sec4Ai4Sec and CoEvolution.

Maciej Żelaszczyk

Samsung AI Center Warsaw

Tutorial 3: [Generative AI] The Noise, the flow, the images.

Saturday / 18 October 9:00 - 13:00

Description:

A brief introduction to flow matching. We will cover the idea and intuition behind flow matching, how it relates to diffusion models, and what training a flow matching model looks like in practice. We will introduce the flow matching setup for continuous data, show how this setup can be modified with the use of optimal transport and close out with a discrete flow matching example.

Prerequisites: Basic knowledge of linear algebra, probability theory, vector fields, physics, and neural networks. Familiarity with Python, PyTorch, Jupyter Notebook.

Biography:

Maciej is a Senior Research Scientist at the Samsung AI Center Warsaw, where he researches topics related to safety and alignment in neural networks. He holds a Ph.D. in computer science from the Warsaw University of Technology with a dissertation on representation learning. Prior to joining Samsung, he worked in a variety of fields, including a stint as Software Engineer at Cardinal Cryptography, a Senior Data Scientist position at Yosh.AI with focus on recommendation systems and as Machine Learning Engineer at Daftcode developing a matchmaking system for mobile games. Apart from that, he also has experience in finance, particularly quantitative investments.

Patryk Wielopolski

Independent Researcher, AI Safety Poland

Taras Kutsyk

Jagiellonian University, AI Safety Poland

Tutorial 4: From Superposition to Sparse Autoencoders: Understanding Neural Feature Representations

Saturday / 18 October 9:00 - 13:00

Description:

Modern neural networks often learn representations that are difficult to interpret, with individual neurons responding to multiple unrelated features - a phenomenon called polysemanticity. This workshop explores the theoretical foundations and practical implications of how neural networks represent features through the lens of Anthropic’s influential “Toy Models of Superposition” paper. Participants will gain hands-on experience with the fundamental concepts of mechanistic interpretability by building and analyzing simple neural networks that demonstrate superposition - the ability of models to represent more features than they have dimensions. Through interactive exercises attendees will train toy models with varying sparsity levels and observe how networks organize features into geometric structures like antipodal pairs and pentagons when forced to compress high-dimensional feature spaces. The workshop concludes with an exploration of Sparse Autoencoders (SAEs) as a promising solution for disentangling features. Participants will implement and train their own SAEs on the toy models, visualizing how dictionary learning techniques can recover interpretable feature representations from seemingly incomprehensible neural activations. This workshop bridges theory and practice, providing both the mathematical intuition behind superposition phenomena and practical tools for investigating real neural network behavior - essential knowledge for anyone interested in AI safety, interpretability research, or understanding the inner workings of modern ML systems.

Prerequisites: Participants should have a basic understanding of deep learning and PyTorch. For this tutorial, we will be using Google Colab, so only a web browser with internet access is required.

Biography:

Patryk Wielopolski is an AI researcher with a Ph.D. in probabilistic modeling and publications at AAAI, ECAI, and TPAMI. He previously served as Solution Innovation Leader at DataWalk, leading the company’s AI research agenda on Knowledge Graphs, Large Language Models, and unstructured data processing. As an active member of Poland’s AI community, he co-founded the genwro.ai research group and contributed to the Polish AI Olympics. At previous ML in PL conferences, he presented on knowledge graphs (2024) and TreeFlow (2023), receiving the Best Contributed Talk Award for the latter. Patryk is currently transitioning his research focus to AI Safety.

Taras Kutsyk is a Ph.D. student specializing in AI Interpretability and Safety. He began his career in machine learning with research on satellite imagery enhancement, before shifting his focus to AI Safety after completing the AI Safety Fundamentals course. Since then, he has completed research internships in mechanistic interpretability of language models, including the MATS program under Neel Nanda and the AI Safety Camp (AISC). His work has led to front-page–promoted blog posts on new insights into Sparse Autoencoders (SAEs). Taras is currently investigating model organisms of misalignment and developing mechanistic approaches to prevent it.

Anna Kozak

Warsaw University of Technology

Katarzyna Woźnica

Warsaw University of Technology, Systems Research Institute, Polish Academy of Sciences

Antoni Zajko

Warsaw University of Technology

Tutorial 5: Introduction to Automated Machine Learning (AutoML)

Saturday / 18 October 9:00 - 13:00

Description:

AutoML aims to democratise machine learning by automating key stages of the model development pipeline - making it accessible to experts and domain specialists. This tutorial offers a comprehensive overview of the AutoML field, including its motivations, history, and technical foundations. We will explore the major components of an AutoML system: algorithm selection, hyperparameter optimisation, meta-learning, and ensemble methods. The human expert’s role in the AutoML pipeline and challenges such as interpretability and monitoring will also be discussed. The practical component of the tutorial includes hands-on experience with popular open-source AutoML frameworks in Python, such as AutoGluon and MLJAR. Participants will build complete AutoML pipelines, compare tools, and understand their strengths and limitations. This tutorial is ideal for ML practitioners, researchers, and data scientists who want to understand how AutoML works under the hood and how to use it effectively in real-world projects.

Prerequisites: Basic knowledge of machine learning concepts (supervised learning, model evaluation, overfitting). Familiarity with Python programming and standard ML libraries (scikit-learn, pandas). Participants should install the following Python-based AutoML libraries in a virtual environment: auto-gluon, mljar. A Jupyter notebook environment (e.g., Anaconda, JupyterLab, or Google Colab) is recommended for hands-on exercises.

Biography:

Anna Kozak is a data scientist with over eight years of experience. She conducts research in Automated Machine Learning at the Warsaw University of Technology, and lectures data visualisation, machine learning, and statistics.

Katarzyna Woźnica holds a PhD in Computer Science with a specialization in Machine Learning. Her work centers on AutoML, hyperparameter optimization, and meta-learning, with a focus on human-in-the-loop methods and Automated Data Science. She has experience applying machine learning methods in medical research and clinical practice.

Antoni Zajko is a PhD student at Warsaw University of Technology with both research and commercial experience in machine learning.

Adam Czajkowski

Tenstorrent

Tutorial 6: Next Generation Hardware for ML - Hands on with Tenstorrent🤘

Saturday / 18 October 9:00 - 13:00

Description:

This workshop offers a guided, hands‑on introduction to running AI workloads on Tenstorrent hardware. Participants will connect to provided environments, experiment with example models, and explore profiling tools. By the end, they’ll have a practical overview of tt-metal and know how to continue developing on Tenstorrent platforms.

Overview:

Walkthrough of Stable Diffusion 3.5, general overview and implementation details surrounding the Tensix architecture. Discussion on multi-core and multi-device challenges, including memory and performance. Followed by an interactive demo.
After the demo the workshop will transition to hands-on segments, guided by Jupyter notebooks. Topics covered will include, but are not limited to:
- An introduction to the Metalium programming model, and a deeper dive into the Tensix architecture.
- Introduction into TT-NN, a Python and C++ open source library of neural network operations.
- MNIST MLP inference guided tutorial.
- TT-NN Visualizer tutorial, an OP optimization and memory usage visualization tool.
- A Tenstorrent model writing introduction guide.
Multiple participants that successfully complete the guided workshop will be awarded Tenstorrent Blackhole and Wormhole cards at the end of the workshop.

Knowledge Prerequisites:

Intermediate knowledge of Python and PyTorch.
Basic understanding of machine learning concepts.
Basic background knowledge of vision classification and LLM models will be useful, but is not required.

Hardware Requirements:

Participants will be required to bring their own laptops.
Connection instructions will be provided to connect to a Tenstorrent VM running on tt-cloud. The VM already has a preconfigured n150 Wormhole card and Jupyter notebook environment.

This tutorial is free of charge but also requires registration. Please register here.

Biography:

Adam Czajkowski leads software engineering for TT-Metal and Developer Experience initiatives across the company’s software stack, including TT-NN, TT-Metal, and supporting tools. In his career, in addition to his work at Tenstorrent, Adam has developed embedded systems and firmware on RISC-V and ARM, as well as conducted post-graduate research on NUMA multi-processor systems, and custom-build operating systems including K42.

Tudor Coman

Adobe

Tutorial 7: Migrating Python AI Prototypes to Cross-Platform Solutions

Saturday / 18 October 14:00 - 18:00

Description:

This tutorial helps active and aspiring developers and engineers understand how to migrate Python-based AI proof-of-concepts into production-ready, cross-platform systems. It focuses on practical strategies for translating NumPy-based computations, model inference, and LLM integrations into Java/Kotlin environments using technologies such as ONNX, Multik, LangChain4j, pgvector and Kotlin coroutines. Learners will gain skills to preserve performance, ensure interoperability, and future-proof AI-driven applications.

Prerequisites:

Proficiency in Python and NumPy
Familiarity with AI model inference and APIs
Some experience with Java (or Kotlin as a bonus)
Software installed: Python 3 with an editor of their choice, Java 21+, IntelliJ IDEA Community Edition (preferably), Gradle, Kotlin

Biography:

Tudor Coman is a Machine Learning Engineer at Adobe with more than seven years of experience. His work spans reinforcement learning, natural language processing, and large language models, alongside deep expertise in developing web services and big data infrastructures that support advanced AI use cases. His current focus includes production-level integration of multi-armed bandits algorithms in experimentation use cases, as well as generative content analysis and suggestions for A/B tests.

Mateusz Olko

University of Warsaw, IDEAS NCBR

Mateusz Gajewski

Poznań University of Technology, IDEAS NCBR

Tutorial 8: Modern Causal Discovery

Saturday / 18 October 14:00 - 18:00

Description:

Complex real-world systems, such as the human body, the global climate, or an economy, can often be decomposed into simpler components that interact with one another, typically in sparse and structured relationships. These interaction patterns, named causal relationships, form the basis for modeling such systems more effectively. The field of causal discovery focuses on recovering these underlying interaction patterns from observational or experimental data and offers a principled solution to this challenge. Contrary to traditional statistical models, causal discovery allows distinguishing correlation from causation, leading to robust conclusions and deepened scientific insight. Causal discovery methods are becoming increasingly central to scientific inquiry across disciplines such as medicine, biology, economics, and climate science, where they help reveal underlying causal mechanisms, inform experimental design, and support more effective decision-making. This tutorial introduces the theoretical foundations and recent advances in causal discovery, with a focus on scalable, modern approaches. We begin with an accessible overview of the theory of causality, emphasizing the distinction between correlation and causation, key assumptions, and common frameworks such as structural causal models and graphical representations. We then review foundational techniques in causal discovery before progressing to state-of-the-art algorithms, that integrate deep learning methods for increased efficiency and flexibility. The tutorial concludes with a discussion of current challenges, open research questions, and practical considerations for applying causal discovery in real-world settings. Attendees will gain a broad yet detailed understanding of the field, equipping them to apply causal discovery techniques to scientific and applied problems across domains.

Prerequisites: Participants should have a working knowledge of undergraduate-level calculus and linear algebra. Familiarity with basic probability and statistics is recommended. Basic Python programming experience is expected, including use of standard scientific libraries such as NumPy and SciPy.

Biography:

Mateusz Olko is a doctoral researcher at the University of Warsaw and IDEAS NCBR. He earned both his Bachelor’s and Master’s degrees in Computer Science from the University of Warsaw, specializing in machine learning. He explores topics in causal machine learning and causal discovery, with a particular interest in how they can be connected to deep learning. More broadly, he is interested in computational reasoning and how learning systems can better capture structure and support decision-making. His research has been recognized at top-tier AI conferences, including NeurIPS and ICML.

Mateusz Gajewski is a PhD student in the Intelligent Algorithms and Data Structures Research Group at Poznań University of Technology and IDEAS NCBR. His primary research interests focus on causality, particularly causal discovery and the application of causal methods in small data settings. He is also interested in explainability, especially approaches involving game theory, as well as the intersection between causality and explainability.

Barbara Klaudel

TheLion.AI / Gdańsk University of Technology

Aleksander Obuchowski

Medalion Technology / TheLion.AI

Tutorial 9: Creating AI Tools for Healthcare

Saturday / 18 October 14:00 - 18:00

Description:

This tutorial explores the development of AI tools tailored for healthcare applications, drawing from real-world projects like the Eskulap ecosystem—a suite of Polish-language models for medical natural language processing. We will dive into key components: large language models (LLMs) for tasks such as medical question-answering and summarization, text ncoders for generating embeddings to support retrieval-augmented systems, Image Encoders for analyzing medical visuals like scans; and automatic speech recognition (ASR) models for transcribing clinical conversations. Participants will learn Parameter-Efficient Fine-Tuning (PEFT) techniques, including Low-Rank Adaptation (LoRA) to mitigate catastrophic forgetting during continual training, Task-Specific Denoising Autoencoders (TSDAE) for robust representation learning, and Contrastive Learning for improving embeddings through positive-negative pair discrimination. We will also cover data acquisition strategies, emphasizing synthetic data generation to address scarcity in medical datasets, alongside cleaning and augmentation methods from sources like scientific publications and Wikipedia. The tutorial balances theoretical foundations with practical implementations, highlighting challenges like data privacy and model efficiency in healthcare. Attendees will gain hands-on experience building modular AI systems, with examples from ongoing research. This session underscores the importance of AI in advancing equitable healthcare, particularly in underrepresented languages and domains.

Prerequisites: Participants should have a basic understanding of machine learning concepts, including neural networks, supervised/unsupervised learning, and familiarity with Python programming. Prior exposure to deep learning frameworks like PyTorch or Hugging Face Transformers is recommended but not mandatory, as the tutorial will include introductory overviews.

Biography:

Barbara Klaudel is a co-founder of an interdisciplinary research group, TheLion.AI, devoted to creating AI-based open-source solutions for healthcare. She leads the project UMIE, which standardizes medical imaging datasets and releases medical imaging encoders. She works as a chief research officer at Medalion Technology. She lectures at the Gdańsk University of Technology. She was awarded Forbes 25 under 25.

Aleksander Obuchowski is a co-founder of an interdisciplinary research group, TheLion.AI, devoted to creating AI-based open-source solutions for healthcare. He leads the project Eskulap, which releases a set of tools for Polish medical NLP (you will learn more about them during our tutorial). He works as a chief technology officer at Medalion Technology. He was awarded Forbes 25 under 25.

Tomasz Steifer

Institute of Fundamental Technological Research, Polish Academy of Sciences

Przemysław Andrzej Wałęga

Queen Mary University of London

Tutorial 10: Navigating the landscape of transformers' expressivity

Saturday / 18 October 14:00 - 18:00

Description:

Why models based on transformers perform great on some tasks and not so well on others? What kinds of problems are easy for transformers and which are hard or even impossible? In recent years, there has been a significant progress in theoretical study of transformers’ expressivity. But also, there has been some misunderstanding. For instance, are transformers really Turing-complete or are they rather weak, basically not able to understand more than regular languages? This tutorial will give an overview of basic results in the area. During these introductory lectures, we will try to understand what is really known about transformers, what are the open questions and especially, what kind of simplifying assumptions are made by some theoreticians of transformers?

Prerequisites: We don’t assume any knowledge except basic understanding of linear algebra and some knowledge from first year undergraduate CS courses. However, the tutorial will be generally easier to follow for people who already have some basic knowledge of the building blocks of the transformer, in particular what is an attention layer and a multi layer perceptron/feedforward neural network.

Biography:

Tomasz Steifer is an adiunkt at the Institute of Fundamental Technological Research of the Polish Academy of Sciences and external collaborator at the Centro Nacional de Inteligencia Artificial (Chile). After obtaining his PhD from the Institute of Computer Science of the Polish Academy of Sciences, he was a postdoc at the Pontificia Universidad Católica de Chile and, through various fellowships, spent time at the Laboratoire Bordelais de Recherche en Informatique, at the University of California, Berkeley and at the University of Bristol. He works on theoretical foundations of machine learning and AI, including learning theory and the expressivity of transformer architecture, as well as on topics in mathematical logic and computational social choice.

Przemysław Wałęga is an Associate Professor (Senior Lecturer) in Queen Mary University of London, Centre for Fundamental Computer Science. He was a Senior Researcher in University of Oxford, Department of Computer Science and received PhD in Logics in the Institute of Philosophy at the University of Warsaw. His research is devoted to designing formal logical languages for AI, studying their computational properties, and developing efficient reasoning algorithms for them. He is especially interested in methods for complex reasoning about time which brings together computer scientists, mathematical logicians, and philosophers.

Mieszko Rutkowski

Allegro

Tutorial 11: Machine Translation in the LLM Era: A Practical Guide

Saturday / 18 October 14:00 - 18:00

Description:

In the LLM era, one might think that the problem of Machine Translation is solved. But it’s certainly not, especially at scale and in domain-specific contexts. This tutorial provides a practical overview of modern MT. I will start with the core concepts like parallel data, system architectures and evaluation methods, and then will move to more advanced topics like quality-aware decoding or applications of reinforcement learning to MT. During the hands-on session, we will implement and evaluate state-of-the-art MT techniques on our own!

Prerequisites: No prior knowledge in Machine Translation is required, but familiarity with basic deep learning concepts is beneficial. Hands-on session will be run in a Jupyter notebook. To actively participate in the tutorial, you need to have Python 3 installed together with libraries transformers and trl.

This tutorial is free of charge but also requires registration. Please register here.

Biography:

Mieszko Rutkowski is a Research Engineer at Allegro, where he contributes to developing Allegro Translation Engine. His interests include Machine Translation, especially with reinforcement learning and multimodal methods. Previously he worked on risk management at UBS and on astrophysics at the Jagiellonian University.

Wojtek Czekalski

CEO at Dialo

Patryk Wojnarowski

Dialo

Adam Cholewiński

Dialo

Tutorial 12: End "It Works On My Machine": A Practical Guide to Nix and uv

Saturday / 18 October 14:00 - 18:00

Description:

Welcome to dependency hell; this workshop is your way out. Learn to replace the chaos of pip, apt, and docker with modern tools used by elite firms like Jane Street and Anduril to build reliable and isolated environments. To achieve this, we’ll first go back to basics, teaching you the Linux fundamentals—how PATH, environment variables, and linking really work—that cause these problems in the first place. Armed with that knowledge, you will learn to use Nix to declaratively define and build perfectly self-contained environments from a single file. The result is true isolation, giving you the power to reliably set up any project and switch between conflicting toolchains instantly. Finally, you’ll turbocharge your Python workflows with the uv package manager and see how these principles scale up, from ensuring dev/prod parity to managing entire, reproducible operating systems with NixOS.

This tutorial is free of charge but also requires registration. Please register here.

Biography:

Wojtek leads development at Dialo. This involves integrating a wide array of technologies, from machine learning models to backend services, making consistent and reproducible environments essential for success.

Patryk is a Software Engineer at Dialo currently focusing on infrastructure. He believes in building elegant, declarative systems as the most effective way to manage complexity and achieve true reproducibility.

Adam is a systems engineer who builds performant and reliable systems mostly using OCaml. He is a strong advocate for using functional programming and declarative tooling to eliminate entire classes of bugs before they ever reach production.

/ Table of Contents

/ Agenda

Day 1: Wednesday / 15 October

Copernicus Science Centre

Wybrzeże Kościuszkowskie 20, 00-390 Warsaw

Student Research Workshop (open event, no ticket required)

PepGlider: Property Regularized VAE for interpretable and controllable peptide design

by Aleksandra Gwiazda (Warsaw University of Technology)

DetoxAI - Python Package for Debiasing Neural Networks

by Łukasz Sztukiewicz (Independent Researcher)

OMENN: One Matrix to Explain Neural Networks

by Mikołaj Janusz (Jagiellonian University)

From Blurry to Sharp: Enabling High-Frequency Detail in Implicit Neural Representations

by Tomasz Dądela (Jagiellonian University)

Semantic Label Reconstruction: How to breach privacy in Federated Learning

by Rafał Malcewicz (Carnegie Mellon University, GHOST Day)

Filling in the Blanks: Redefining Inpainting with a Novel Geometrical Latent Space VAE Architecture

by Nikodem Świerkowski (Wrocław University of Science and Technology)

GraphaFold: Graph Neural Network for predicting non-canonical RNA base pairing

by Paulina Hładki (Poznan University of Technology)

Enhancing Chemical Explainability Through Counterfactual Masking

by Łukasz Janisiów (Jagiellonian University)

Registration

(Also open after 12:00)

Opening remarks

Opening remarks

Discussion Panel 1: PL in ML

(conducted in Polish)

Invited Talk 1: Causal representation learning in temporal settings with actions

by Sara Magliacane (University of Amsterdam)

Invited Talk 2: Architecture for open models

by Shreya Pathak (Google DeepMind)

Lunch

Sponsor Talk: Sequential Representation Learning for Real-Time Bidding

by Mateusz Błajda and Maciej Zdanowicz (RTB House)

Sponsor Talk: Evolving Search and Recommendations at a Leading E-commerce Platform

by Aleksandra Chrabrowa and Mateusz Marzec (Allegro)

Partner Talk: Phoenix from the ashes - IDEAS Research Institute

by Tomasz Trzciński (IDEAS Research Institute)

Invited Talk 3: Technical AI Safety & Security: From Alignment to Agentic Systems

by Adel Bibi (University of Oxford / Kellog College / Softserve)

Invited Talk 4: Learning to see the hidden world: A perspective on causal representations

by Francesco Locatello (ISTA)

Conference Party (powered by Dialo)

Bolek Pub & Restaurant, al. Niepodległości 211, 02-086 Warszawa

Remember to take your badges and ID cards with you, they are required to enter

Day 2: Thursday / 16 October

Copernicus Science Centre

Wybrzeże Kościuszkowskie 20, 00-390 Warsaw

Registration + Breakfast

(Also open after 09:30)

Contributed Talks Session 1

Learning Dynamic Segmentation and Compression of Sequences in Transformer LLMs

by Adrian Łańcucki (NVIDIA)

State-of-the-Art Document AI on a Single 24GB GPU

by Łukasz Borchmann (Snowflake)

Scaling Fine-Grained MoE Beyond 50B Parameters: Empirical Evaluation and Practical Insights

by Jakub Krajewski (University of Warsaw)

Contributed Talks Session 2

FlySearch: Exploring how vision-language models explore

by Adam Pardyl (Jagiellonian University; IDEAS NCBR)

How Good Are Open-Source Models for Robot Learning?

by Gracjan Góral (University of Warsaw)

Shaping Robotic Actions with Fourier Flow Matching

by Mateusz Wyszyński (University of Warsaw)

Contributed Talks Session 3

Cache Me If You Can: Reducing Model Size and KV Cache Traffic for Faster LLM Inference

by Konrad Staniszewski (NVIDIA, University of Warsaw)

Neural self-supervised audio representation for SpeechLLM: neural audio codecs for Polish language

by Paweł Cyrta (AGH Cyfronet / Stenograf.io)

Entity Tracking as a Microcosm of Semantic Abilities in LLMs and Humans

by Karolina Drożdż (IDEAS Research Institute)

Invited Talk 5: Training LLMs: Do We Understand Our Optimizers?

by Antonio Orvieto (ELLIS Institute Tübingen)

Invited Talk 6: Beyond the screen: capturing, understanding and generating 3D scenes

by Federico Tombari (Google / Technical University of Munich)

Invited Talk 7: Reliable and Sustainable AI: From Mathematical Foundations to Next Generation AI Computing

by Gitta Kutyniok (DLR / University of Tromsø / LMU Munich)

Poster Session 1 + Coffee

with posters: 1-30