Biography:
Pierre Ménard is a Research Scientist at Meta, where he works on large language model agents. His research focuses on reinforcement learning, with particular interests in multi-agent systems, unsupervised learning, and reinforcement learning from human feedback. Before joining Meta, he was a postdoctoral researcher, most recently at ENS Lyon. He earned his PhD in machine learning from Université Toulouse III – Paul Sabatier, where he studied exploration-exploitation trade-offs in sequential decision-making problems.

Abstract:
We introduce Meta Agents Research Environments (ARE), a research platform for scalable creation of environments, integration of synthetic or real applications, and execution of agentic orchestrations. ARE provides simple abstractions to build complex and diverse environments, each with their own rules, tools, content, and verifiers, helping to bridge the gap between model development and real-world deployment. We also propose Gaia2, a benchmark built in ARE and designed to measure general agent capabilities. Beyond search and execution, Gaia2 requires agents to handle ambiguities and noise, adapt to dynamic environments, collaborate with other agents, and operate under temporal constraints. Unlike prior benchmarks, Gaia2 runs asynchronously, surfacing new failure modes that are invisible in static settings. Our experiments show that no system dominates across the intelligence spectrum: stronger reasoning often comes at the cost of efficiency, and budget scaling curves plateau, highlighting the need for new architectures and adaptive compute strategies. Perhaps more importantly, ARE abstractions enable continuous extension of Gaia2 to other environments, empowering the community to rapidly create new benchmarks tailored to their domains. In AI's second half, progress increasingly depends on defining meaningful tasks and robust evaluations to drive frontier capabilities forward.

Biography:
Rebekka leads the Relational Machine Learning Group at the Helmholtz Center CISPA since 2021. Their research on sparse deep learning is supported by an ERC starting grant since December 2023 and Apple Research since August 2025. From 2019-2021, she was a PostDoc at the Biostatistics Department of the Harvard T.H. Chan School of Public Health working with John Quackenbush, from 2017-2018 at the Institute for Machine Learning at ETH Zurich with Joachim Buhmann, and from 2016-2017 at the Chair of Systems Design at ETH Zurich with Frank Schweitzer. Her PhD research at the ETH Risk Center was supervised by Frank Schweitzer and co-supervised by Hans J. Herrmann from 2013-2016. Her thesis on systemic risk won the Zurich Dissertation Prize and our work on international maize trade received the CSF Best Contribution Award. She studied Mathematics and Physics at TU Darmstadt.

Abstract:
Deep learning continues to achieve impressive breakthroughs across disciplines but relies on increasingly large neural network models that are trained on massive data sets. Their development inflicts costs that are only affordable by a few labs and prevent global participation in the creation of related technologies. In this talk, we will ask the question whether it really has to be like this and discuss some of the major challenges that limit the success of deep learning at smaller scales. Last but not least, we will also provide an overview over our most promising solution strategies.

Biography:
After graduating from École Polytechnique, Louis worked on multimodal generative models at the University of Cambridge. He is now the CEO of Retab, a platform used by AI and operations teams, from fast-growing startups to Fortune 500 companies, to automate their hardest document workflows.

Abstract:
Most back-office work is built around processing documents at scale. Insurance companies review thousands of claim packets per day. Logistics firms reconcile shipping documents across millions of orders per year. Underwriting teams extract the same fields and check the same rules - thousands of times. These aren't one-off analytical tasks. They're high-volume, repetitive workflows that companies staff entire teams to run manually. Vibe coding a "document agent" demo with OCR + Gemini is easy. Scaling it to production with minimal human review is hard. When your extraction accuracy is 90% instead of 99%, every output needs human verification - and review time approaches the time of doing it yourself. This was the fundamental limitation of traditional IDP, and it hasn't gone away just because the models got better. This talk will cover the latest on building multi-stage VLM pipelines that scale to millions of documents while minimizing human review: schema-driven structured extraction, context management for large files, k-LLM consensus for uncertainty quantification, and agentic optimization loops that maximize F1 scores. The session ends with a live demo.

Biography:
Dr.Marina Esteban-Medina is a postdoctoral researcher at the ETH Zürich AI Center and ETH Zürich’s Department of Biosystems Science and Engineering (Switzerland). She works at the intersection of AI and translational biomedicine, developing machine-learning and statistical models to understand tumor heterogeneity and treatment response, with a particular focus on pediatric cancer. She earned a degree in Biotechnology and an MSc in Computational Biology from Universidad Carlos III de Madrid, followed by a PhD in Computer Science from the University of Córdoba. During her doctoral work in a Computational Medicine platform, she built predictive models that integrated clinical and genomic data from the Andalusian healthcare system with public resources to identify therapeutic targets in cancer and rare diseases, collaborating with the Francis Crick Institute (London) and Institut Curie (Paris). Her research has been supported by competitive international funding, including an EMBO Scientific Exchange Grant and an ETH AI Center Postdoctoral Fellowship. She contributes to international initiatives such as the Disease Map and several ELIXIR communities as well as the Immune Digital Twins. Her interests lie in bridging robust AI methodology with clinically actionable insights.

Abstract:
Predicting tumor drug response is a fundamental challenge in precision oncology, complicated by cellular heterogeneity, where minor resistant sub-populations drive clinical relapse. While single-cell transcriptomics offers a high-resolution view of these mixtures, a significant domain gap remains: most drug-response labels originate from bulk cell-line screens, whereas clinical deployment targets single-cell patient data. In this talk, I will explore how to bridge this gap through biologically grounded domain adaptation. I will examine why standard alignment-focused deep transfer methods often fail when they overlook the complex biological realities of the transfer problem. Robust translation requires moving beyond predicting a single sensitivity score toward modeling response as a structured change in cellular state. In this direction, recent generative frameworks have opened promising paths toward a comprehensive modeling of these cellular responses. Finally, I will highlight the importance of uncertainty estimation as a guardrail, ensuring that predictions in new single-cell datasets are safely bounded by the limits of the model's training knowledge.

Biography:
Joann is a PhD researcher in Computer Science at the Institute of Computational Perception, Johannes Kepler University. She holds a Bachelor's degree (BM) in Music Performance (Violin) from The University of Texas at Austin and a Master’s degree (MSc) in Music Technology from the Georgia Institute of Technology. She previously worked as a Research Assistant at the Music and AI Lab at Academia Sinica in Taipei, Taiwan. Her research focuses on Music Emotion Recognition, with an emphasis on understanding how emotional expression varies across performances, which lies at the intersection of audio signal processing, computational music analysis, and music cognition. With a background in music performance, she continues to engage in community outreach through music. She is also a member of the Philharmonia Moments Musicaux in Taiwan, where she previously served as Principal Second Violin and remains involved when visiting.

Abstract:
Music Emotion Recognition (MER) aims to model and understand how music conveys emotion, and lies within the broader field of Music Information Retrieval (MIR). Despite recent advances in machine learning, MER remains challenging due to issues such as limited data and generalization across datasets. In this talk, she will provide a high-level introduction to MIR and discuss recent research on MER utilizing audio and symbolic modalities to model emotional content. She will present findings on the generalization gap across datasets and explore methods to mitigate this gap. She will also discuss ongoing work investigating how performance differences between musicians influence perceived emotion, and briefly highlight potential applications such as therapeutic uses of emotion-aware music technologies.

Biography:
Ondřej Dušek is an Assistant Professor at Charles University in Prague, working on natural language generation and human-computer dialogue. His research focuses on generative language models including large language models, mostly applied to the data-to-text and dialogue response generation tasks. He is specifically interested in evaluating the quality of generated content, specifically its semantic accuracy. After obtaining his PhD in Prague, Ondřej spent 2 years as a postdoc at Heriot-Watt University in Edinburgh. Back in Prague, he is currently the PI of an ERC Starting Grant which aims to produce fluent, accurate and explainable natural language generation systems.

Biography:
I am an assistant professor in Natural Language Processing at the University of Edinburgh and an affiliated lecturer at the University of Cambridge. Over the past year, I was a visiting professor at NVIDIA. My research focuses primarily on efficient and modular architectures for foundation models, especially with respect to adaptive memory and end-to-end tokenization. Previously, I was a visiting postdoctoral scholar at Stanford University and a postdoctoral fellow at Mila Montreal and Mc Gill University. In 2021, I obtained a PhD from the University of Cambridge. My research has been featured the Economist and Scientific American, among others. I received a Google Research Faculty Award and several awards (Highlight Awards at ACL 2025 and Best Paper Awards at EMNLP 2021 and Repl4NLP 2019). I am recipient of an ERC Starting Grant and a £2M ARIA grant. I am a Scholar of the European Lab for Learning and Intelligent Systems (ELLIS) and part of the TACL journal editorial team.

Biography:
Ivan Cimrák is a Professor at University of Žilina (Slovakia), where he works at the intersection of artificial intelligence and biomedical engineering. His recent research focuses on neural networks for medical imaging, with an emphasis on mammography—including analyses of mammography datasets for deep learning and methods that leverage longitudinal screening information to support earlier and more reliable detection. In collaboration with clinical partners, he also develops AI-assisted workflows for digital pathology, such as computer-supported scoring of proliferation markers (e.g., Ki67) from routine histology slides. Alongside AI, he has a strong background in computational modelling of blood flow and cell mechanics, and he has contributed widely used open-source implementations for simulating deformable cells in fluid environments. His work has been recognized by multiple awards, including the Slovak “Cena za vedu a techniku” (Science and Technology Award) 2024 in the category Personality of Science and Technology, and he was a finalist of the ESET Science Award 2024 (Outstanding Academic).

Abstract:
If AI can match or even outperform radiologists in benchmark studies, why is it still not running screening programs? This talk looks at mammography as a stress test for applied ML: low signal-to-noise ratio, rare positives, subtle findings, multimodal context, and high-stakes decisions. I will show that success in retrospective studies is only part of the story. The real challenge is bridging the gap between strong models and trustworthy clinical deployment.

Biography:
I am a PhD candidate in Computer Science at AGH University of Krakow, and the leader of Machine Learning and Chemoinformatics Lab (MLCIL). My research concerns fair evaluation, graph representation learning, graph classification, chemoinformatics, and molecular property prediction. I'm also interested in time series, NLP, and MLOps, and I'm also teaching all of those things at AGH. I also work at Placewise as Data Science Engineer, and at MatGen as Senior Chemoinformatician. Beside my professional work, I train Historical European Martial Arts (HEMA) with messer and longsword, and like reading and tabletop RPGs.

Abstract:
Carcinogenicity - ability of molecules to cause cancer - is the critical regulatory endpoint for both drugs in medicinal chemistry and pesticides in agrochemistry. Testing it for novel compound candidates is long, hard, and expensive. At the same time, cancer safety is of paramount importance - glyphosate cancer-related court cases alone resulted in almost $10 billion in settlements so far. Over the years, many methods have been proposed, spanning feature engineering, graph neural networks (GNNs), and NLP-inspired transformers for molecules. However, all of them have been tested almost exclusively on medicinal chemistry data. What happens if we try to apply them to pesticides? As we will show, all of those models break down. Even slight data distribution shift causes existing models to fall in performance. While pesticides are similar to small molecule drugs, no current solutions have acceptable quality to predict their carcinogenicity. We will cover data analysis and machine learning approaches in chemoinformatics that allow us to quantitatively evaluate those issues, and how we can alleviate those problems with data augmentation, transfer learning, and out-of-distribution learning methods.

Biography:
Human-centric AI researcher at the IDEAS Research Institute, working in the Phenomenology and Computational Psychiatry group. She holds a Research Master’s degree in Neuroscience and Artificial Intelligence from the University of Amsterdam. Her work bridges cognitive science, psychology, and deep learning to study AI’s (mis)alignment with human cognition. She evaluates LLMs’ functional capacities for introspection, metacognition, and situation modelling in relation to human cognitive abilities. In parallel, she uses NLP to analyse and model human lived experience, particularly in complex clinical contexts.

Abstract:
Growing reliance on LLMs for psychiatric self-assessment raises questions about their ability to interpret qualitative patient narratives. We present the first direct comparison between state-of-the-art LLMs and mental health professionals in diagnosing Borderline (BPD) and Narcissistic (NPD) Personality Disorders utilizing Polish-language first-person autobiographical accounts. We show that the top-performing Gemini Pro models surpassed human professionals in overall diagnostic accuracy by 21.91 percentage points (65.48% vs. 43.57%). While both models and human experts excelled at identifying BPD (F1 = 83.4 & F1 = 80.0, respectively), models severely underdiagnosed NPD (F1 = 6.7 vs. 50.0), showing a reluctance toward the value-laden term "narcissism." Qualitatively, models provided confident, elaborate justifications focused on patterns and formal categories, while human experts remained concise and cautious, emphasizing the patient's sense of self and temporal experience. Our findings demonstrate that while LLMs are highly competent at interpreting complex first-person clinical data, they remain subject to critical reliability and bias issues.

Biography:
He is a computer science engineer and artificial intelligence master's graduate, passionate about coding, automation, and AI agents. He is dedicated to knowledge sharing through talks, articles, and video courses. His journey in AI began with DeepMind's AlphaZero chess system, reflecting his love for the game. His interests later evolved to include computer vision and autonomous vehicles, and eventually expanded to large language models and agentic systems. With professional experience in HRTech and MedTech startups focusing on process automation, he is currently developing MisterCar — an agentic framework for researchers and developers. When not immersed in AI, he enjoys watching and playing chess, his favorite games including Stronghold and S.T.A.L.K.E.R., and listening to epic soundtrack music from movies and games.

Abstract:
MisterCar is a configuration-driven framework for building AI agents that interact with virtual environments. This talk demonstrates practical applications through four complete imitation learning examples. I'll showcase how MisterCar enables rapid experimentation in computer vision research by transforming video games into AI testbeds. Through live demonstrations of bots trained for Space Waves, Pacman, ElectricMan, and Euro Truck Simulator 2, you'll see how the same framework adapts to different ML tasks — from binary classification for timing-based games, through multiclass navigation, to multilabel combat systems and regression for continuous control. The presentation focuses on the complete development workflow: data collection from gameplay demonstrations, exploratory data analysis, neural network training with PyTorch Lightning, and real-time agent deployment. Each example illustrates how MisterCar's modular architecture — sensors, transforms, planners, and executors — composes into production-ready agents entirely through JSON configuration, eliminating the need for custom training loops or inference pipelines. Key takeaways include understanding when to apply different classification approaches, handling temporal context in visual decision-making, and leveraging configuration-driven development for rapid AI prototyping. Whether you're a CV researcher seeking experimental environments or a developer interested in practical imitation learning applications, this talk demonstrates how sophisticated AI agents can be built through composition rather than custom programming.

Biography:
Mateusz is a PhD candidate at the Faculty of Computer Science, AGH University of Krakow, and a member of the AGH ML and Chemoinformatics Group. His research focuses on molecular representation learning and fair evaluation in cheminformatics. In his most recent study, he benchmarked 25 foundation models for molecular property prediction across 25 datasets using hierarchical Bayesian statistical testing. Beyond academia, Mateusz works as a Senior Data Science Engineer at Sabre Labs R&D, developing machine learning models for the airline and hospitality industries.

Abstract:
When comparing machine learning models, statistical testing helps us understand whether the observed differences are real, or just caused by random noise. Most researchers rely on null hypothesis significance testing (NHST) when analyzing their results, however many authors have raised concerns about it. NHST cannot describe the probability of one model being better than another, it cannot confirm whether two models perform practically equivalently, and a statistically significant result does not always mean the difference is meaningful in practice. Bayesian statistical testing offers a more informative alternative. It gives you the actual probability that one algorithm outperforms another, lets you formally claim that two models are practically equivalent, and makes uncertainty explicit, all without setting an arbitrary p-value threshold. In this talk, we will go through the differences between the two approaches, how Bayesian testing handles multi-model, multi-dataset evaluation scenarios, and show how you can add Bayesian testing to your own workflow with just a few lines of Python code.

Biography:
I'm a Data Scientist and the Manager of the Risk Machine Learning Team at AllegroPay , where I specialize in credit risk and fraud detection within the financial sector. Before moving into management, I held senior data science roles at institutions like PKO BP and Hexaware / AXA GO. During that time, my work heavily focused on building robust machine learning models using Gradient Boosting and Neural Networks. My technical foundation includes a Master of Computer Science (and Automatic Control and Robotics) from the Warsaw University of Technologies and an MBA in Innovation and Data Analysis. I'm highly passionate about the ethical implications of AI, which is why I am currently pursuing Doctoral Studies in Economics at the Warsaw School of Economics. There, my thesis research focuses specifically on fairness in credit risk modeling.

Abstract:
Machine learning models consume raw, tabular data, but business stakeholders think in complex concepts and relationships. This mismatch creates a "Semantic Gap," which quietly builds up "Semantic Debt" across ML systems over time. In this presentation, I will walk you through how to bridge that gap using Knowledge Graphs. Moving beyond standard data dictionaries, I'll show you the practical mechanics of mapping ML features to structured semantic layers, using frameworks like FIBO as a reference point. We will look at actual use cases where this semantic mapping directly improves the ML lifecycle. You will see how it guides smarter feature selection, highlights blind spots in feature engineering, and dramatically improves Explainable AI (XAI) by aggregating raw variables into intuitive, human-readable concepts. Finally, we'll briefly cover how structuring your data this way provides the necessary foundation for building reliable, autonomous AI agents.

Biography:
I have over 18 years of experience in the IT industry, specializing in data analysis, data processing, and data architecture. For more than 8 years, I have worked in Data Science, gaining strong expertise across corporate finance, capital markets, pharmaceuticals, and medical devices. Certified Microsoft Data Science Associate and Microsoft Data Analyst Associate. I Have over 10 years of mentoring experience, with a particular focus on Data Science, supporting specialists in their professional growth and the practical application of data-driven solutions.

Abstract:
What if we could see deeper, clearer, and faster beneath the Earth’s surface? We explore how modern AI techniques—particularly advanced U-Net architectures—are transforming seismic data interpretation through superresolution. Geophysical data is often noisy, low-resolution, and expensive to process. By applying attention mechanisms, improved loss functions (MSE, MS-SSIM, Gradient and Frequency losses), and transitioning from 2D to 3D models, we significantly enhance image quality, improve signal-to-noise ratio, and preserve critical geological structures such as faults and reflectors. Beyond theory, this session focuses on practical engineering decisions: mixed precision training, kernel optimization, dynamic signal scaling, and cost-efficient deployment. We also discuss evaluation strategies combining quantitative metrics with expert-driven geological validation. This talk is ideal for AI practitioners and domain experts who want to see how deep learning moves from research papers into high-impact industrial applications—delivering sharper insights, faster analysis, and measurable business value in geophysics.

Biography:
Maciej Szymkowski, PhD, is a Senior Machine Learning Engineer at Future Processing, specializing in the intersection of advanced research and industrial AI application. With a career spanning institutions such as Bialystok University of Technology (2018-2026), Warsaw University of Technology (2021-2022), and AGH University in Cracow (2021-2022), he brings a deep academic foundation to complex engineering challenges. Previously, Maciej served as the Head of AI Development at the Łukasiewicz Research Network – Poznan Institute of Technology (2022-2024). His extensive professional portfolio includes high-impact roles at SoftServe, Hemolens Diagnostics, and Transition Technologies, where he transitioned cutting-edge algorithms into production-ready solutions. His specialization is Computer Vision with a focus on medical diagnostics and transport systems (aviation and automotive sector). Moreover, he is an author or co-author of over 45 research papers published in JCR journals and international conference proceedings. When not advancing the field of AI, Maciej is a dedicated sports enthusiast, closely following Real Madrid, the New York Knicks, and the Pittsburgh Penguins.

Abstract:
In 2016, Professor Geoffrey Hinton famously predicted that AI would soon replace radiologists. A decade later, the human expert remains central to clinical practice, yet the technological landscape has been fundamentally reshaped. This presentation provides a critical evaluation of this "prophecy" by contrasting classical methodologies with cutting-edge AI breakthroughs. The talk is structured into two technical phases – in the first, I will concentrate on classical foundations - traditional signal and image processing pipelines (based on algorithms like filtering and thinning) are utilized for initial pathological change detection in anatomical scans (e.g., liver radiology scans). The second part of the talk will concentrate on novel generative frontier. I will explore the transition to modern architectures, evaluating object detection models (YOLO) and specialized foundation models like MedSAM for zero-shot medical segmentation and changes observation. Furthermore, we investigate the reasoning capabilities of Vision-Language Models (VLMs), including Gemini, GPT, Qwen, and InternVL, in interpreting complex diagnostic data. By synthesizing classical processing, deep learning pipelines, and multimodal reasoning, this talk concludes with a data-driven assessment of the current state of the art. Is it finally time to replace the radiologist, or has AI instead created the ultimate diagnostic collaborator?