GMUM

26/05/2026

🎤 It’s GMUM Seminar time again! We have three speakers this week! 🤩

➡️who? Dmytro Polishchuk
➡️where? Wydział Matematyki i Informatyki UJ
➡️when? Thursday (28 Thursday 2026) at 14.15
➡️Talk: “QModel: Time-Aware GitHub Mining Framework for Empirical Software Quality Studies”

Abstract:
Quality-oriented mining software repositories studies require more than access to commits, issues, pull requests, and CI logs. They require a reproducible way to represent how technical, social, temporal, and defect-related evidence is linked across repository history. Existing MSR infrastructures provide important support for repository mining, but they do not directly expose a project-level, time-aware empirical model in which cross-artifact quality evidence can be queried, inspected, and reused across alternative study designs.

This paper presents QModel, an emerging infrastructure for reproducible GitHub mining in empirical software quality studies. QModel stores Git, GitHub, CI, timeline, reaction, graph, churn, and SZZ-style candidate-defect evidence in a project-centric relational schema. Its SQL-based compilation layer allows researchers to define feature-target datasets over arbitrary analysis units, including pull requests, issues, commits, files, time windows, and custom cross-artifact objects.

We evaluate QModel on ansible/ansible and facebook/react, mining 76,475 commits, 70,509 pull requests, 45,908 issues, 2,174,020 timeline events, 315,235 file-change records, and 446,161 CI records. The early results show that QModel can materialize analysis-ready datasets with computable target, graph, churn, and provenance features, while also making evidence gaps explicit when historical links, CI records, or fixing evidence are incomplete. These results support QModel as an empirical infrastructure for making quality-oriented repository operationalizations explicit, reusable, and systematically evaluable.

➡️who? Turhan Can Kargin
➡️where? Wydział Matematyki i Informatyki UJ
➡️when? Thursday (28 Thursday 2026) at 14.15
➡️Talk: “SpaRRTa: A Synthetic Benchmark for Evaluating Spatial Intelligence in Visual Foundation Models”
➡️Project page: https://sparrta.gmum.net/

Abstract:
Visual Foundation Models (VFMs), such as DINO and CLIP, excel in semantic understanding of images but exhibit limited spatial reasoning capabilities, which limits their applicability to embodied systems. As a result, recent work incorporates some 3D tasks (such as depth estimation) into VFM training. However, VFM performance remains inconsistent across other spatial tasks, raising the question of whether these models truly possess spatial awareness or overfit to specific 3D objectives. To address this question, we introduce the Spatial Relation Recognition Task (SpaRRTa) benchmark, which evaluates the ability of VFMs to identify static directional relations between visible objects. Unlike traditional 3D objectives that focus on precise metric prediction (e.g., surface normal estimation), SpaRRTa probes a fundamental capability underpinning more advanced forms of human-like spatial understanding. SpaRRTa generates an arbitrary number of photorealistic images with diverse scenes and fully controllable object arrangements, along with freely accessible spatial annotations. Evaluating a range of state-of-the-art VFMs, we reveal significant disparities in their ability to encode directional spatial relations. Through our analysis, we provide insights into the mechanisms that support or hinder this form of spatial awareness in modern VFMs. We hope that SpaRRTa will serve as a useful tool for guiding the development of future spatially aware visual models.

➡️who? Piotr Helm
➡️where? Wydział Matematyki i Informatyki UJ
➡️when? Thursday (28 Thursday 2026) at 14.15
➡️Talk: “SHIELD: Secure Hypernetworks for Incremental Expansion Learning Defense”

Abstract:
Continual learning under adversarial conditions remains an open problem, as existing methods often compromise either robustness, scalability, or both. We propose a novel framework that integrates Interval Bound Propagation (IBP) with a hypernetwork-based architecture to enable certifiably robust continual learning across sequential tasks. Our method, SHIELD, generates task-specific model parameters via a shared hypernetwork conditioned solely on compact task embeddings, eliminating the need for replay buffers or full model copies and enabling efficient over time. To further enhance robustness, we introduce Interval MixUp, a novel training strategy that blends virtual examples represented as ℓ∞ balls centered around MixUp points. Leveraging interval arithmetic, this technique guarantees certified robustness while mitigating the wrapping effect, resulting in smoother decision boundaries. We evaluate SHIELD under strong white-box adversarial attacks, including PGD and AutoAttack, across multiple benchmarks. It consistently outperforms existing robust continual learning methods, achieving state-of-the-art average accuracy while maintaining both scalability and certification. These results represent a significant step toward practical and theoretically grounded continual learning in adversarial settings.

22/05/2026

🔴 Komitet Informatyki Polska Akademia Nauk wydał krytyczną opinię o projekcie „Polityki rozwoju sztucznej inteligencji w Polsce do 2030 roku".

⚠️ Najmocniejszy zarzut: dokument nie wskazuje żadnych dedykowanych środków finansowych. Ocena Skutków Regulacji wprost stwierdza brak skutków finansowych dla sektora publicznego i brak nowych programów wydatkowych.

Wniosek Komitetu jest jednoznaczny: wstrzymać procedowanie dokumentu i skierować go do ponownego opracowania.

📄 Pełna treść opinii:
https://ki.pan.pl/app/uploads/2026/05/Uchwala_2026_07_zalacznik.pdf

Dziękujemy przewodniczącemu Sekcji Sztucznej Inteligencji przy Komitecie Informatyki PAN, prof. Jerzemu Stefanowskiemu, za udostępnienie uchwały. 🙏



Uniwersytet Jagielloński
GMUM

20/05/2026

🎤 Attend this week’s GMUM seminar and uncover something new.

➡️ who? Anastasiya Pechko
➡️ where? Wydział Matematyki i Informatyki UJ
➡️ when? Thursday (21 May 2026) at 14.15
➡️ Talk: “AEGIS: Preserving Privacy of 3D Facial Avatars with Adversarial Perturbations”

Abstract:
The growing adoption of photorealistic 3D facial avatars, particularly those utilizing efficient 3D Gaussian Splatting representations, introduces new risks of online identity theft, especially in systems that rely on biometric authentication. While effective adversarial masking methods have been developed for 2D images, a significant gap remains in achieving robust, viewpoint-consistent identity protection for dynamic 3D avatars. To address this, we present AEGIS, the first privacy-preserving identity masking framework for 3D Gaussian Avatars that preserves non-identity facial attributes while concealing biometric identity. Our method aims to conceal identity-related facial features while preserving the avatar's perceptual realism and animation usability. AEGIS applies adversarial perturbations to the Gaussian color coefficients, guided by a pre-trained face verification network, ensuring consistent protection across multiple viewpoints without retraining or modifying the avatar's geometry. AEGIS achieves complete de-identification, reducing face retrieval and verification accuracy to 0%, while maintaining high perceptual quality. It also preserves key facial attributes such as age, race, gender, and emotion, demonstrating strong privacy protection with minimal visual distortion. We further propose a region-adaptive variant (AEGIS RA) that reduces visual artifacts by assigning FLAME-guided per-region perturbation budgets while maintaining strong de-identification.

Be there to see it live! 🔥

19/05/2026

📢 Polecamy rozmowę naszych badaczy – prof. Bartosz Zieliński i Adama Wróbla – w Radiu Kraków o metodzie DAVE i tym, jak wyjaśniać decyzje modeli AI. 🎙️

Uniwersytet Jagielloński
Wydział Matematyki i Informatyki UJ
Jagiellońskie Centrum Sztucznej Inteligencji

– „Jesteśmy w stanie zobaczyć, jaki był powód decyzji modelu” – tłumaczył w Radiu Kraków Adam Wróbel z Uniwersytetu Jagiellońskiego. Badacze rozwijają narzędzie, które ma pomóc lepiej rozumieć działanie chatbotów i systemów rozpoznających obrazy.

07/05/2026

🎤 Get ready for an inspiring presentation at this week’s GMUM Seminar.

➡️ who? Dawid Zapolski
➡️ where? Wydział Matematyki i Informatyki UJ
➡️ when? Thursday (7 may 2026) at 14.15
➡️ Talk: “Sampling statistical systems on the lattice efficiently with transformers”

Abstract:
Sampling from high-dimensional discrete Boltzmann distributions is the central computational problem of statistical physics. It is traditionally tackled by Markov Chain Monte Carlo, which suffers from long autocorrelation times near phase transitions. Variational Autoregressive Networks (VAN) take a different route and approach the problem as a generative modeling task: an autoregressive model is trained by minimizing the reverse Kullback–Leibler divergence to the target distribution, and the resulting exact likelihoods can then be used for unbiased reweighting. Notably, training requires no external data, only samples drawn from the model itself. Transformers have so far been considered too costly for this setting, since generating one spin per step leads to a prohibitively long context. We show that the opposite is true. By grouping spins into small patches that are emitted as single tokens, in the spirit of vision transformers, and by adding a physics-informed prior to the output logits based on local energy, even a small transformer can become a state-of-the-art neural sampler. I will briefly introduce the methods and present our results for the 2D Ising model and the Edwards–Anderson spin glass, including the largest systems for which neural-network samplers have been trained to date.

Be there as it happens! 🔥

06/05/2026

🏆 Top 2,2% prac na świecie! Spośród niemal 24 tysięcy zgłoszeń na ICML 2026 tylko niewielki odsetek otrzymuje wyróżnienie spotlight. W tym elitarnym gronie znalazła się praca zespołu badaczy z naszego uniwersytetu.

🧠 Naukowcy z Jagiellońskie Centrum Sztucznej Inteligencjii i Grupy Metod Uczenia Maszynowego (GMUM), we współpracy z badaczami z Niemiec, opracowali metodę DAVE, która pozwala zajrzeć „do środka” modeli sztucznej inteligencji i zrozumieć, na jakiej podstawie podejmują one swoje decyzje.

💡 Dotychczasowe metody wyjaśniania ich decyzji dawały niestabilne, niewyraźne wyniki. DAVE rozwiązuje ten problem przez matematycznie uzasadnione „rozkładanie” tego, jak model przetwarza obraz.

29/04/2026

🎤 Looking ahead to the next GMUM seminar? Two impressive talks are scheduled for this week!

➡️ who? Artur Kasymov
➡️ where? Wydział Matematyki i Informatyki UJ
➡️ when? Thursday (30 April 2026) at 14.15
➡️ Talk: “Diffusion Models Inference Acceleration”

Abstract:
Diffusion models have become the leading approach for image and video generation, but their iterative denoising process makes inference slow and expensive. This talk briefly covers general machine learning techniques for inference acceleration, then deep dives into methods specific to diffusion models — including step reduction, attention optimization, and caching-based approaches that exploit redundancy across denoising steps. In the final part, we present our current work in progress on a novel spatially-adaptive caching method for diffusion transformers.

➡️ who? Witold Wydmański
➡️ where? Wydział Matematyki i Informatyki UJ
➡️ when? Thursday (30 April 2026) at 14.15
➡️ Talk: “Extending Context in Transformer Models: From Positional Extrapolation to Subquadratic Attention”

Abstract:
Scaling context length in transformer models exposes fundamental limitations of self-attention, including its quadratic complexity and dependence on positional representations that often fail to extrapolate. We will analyze positional extrapolation strategies, focusing on how models generalize beyond the training regime. We will then examine architectural modifications that reduce the computational burden of attention, including structured sparsity and locality constraints.

Tune in live and contribute to the conversation!🔥

23/04/2026

🎤 Join us at the GMUM seminar this week and broaden your perspective.

➡️ who? Adam Pardyl
➡️ where? Wydział Matematyki i Informatyki UJ
➡️ when? Thursday (23 April 2026) at 14.15
➡️ Talk: “Object Navigation with Vision-Language Models ”

Abstract:
Real-world environments are inherently noisy and unstructured, necessitating active, goal-oriented exploration to identify relevant information. Although Vision-Language Models (VLMs) demonstrate strong performance in many tasks, their effectiveness in open-ended settings requires further improvements. As we previously shown in our work "FlySearch", state-of-the-art VLMs significantly underperform humans in tasks requiring systematic search. In this presentation, we show our ongoing work on training reasoning vision-language models for active exploration via UAV-based object search in 3D environments. The talk will present our research findings, as well as serve as an introduction to LLM post-training for multimodal reasoning problems.

Don’t miss it! 🔥

15/04/2026

🎤 Be part of this week’s GMUM seminar and discover something new.

➡️ who? prof. Jacek Tabor
➡️ where? Wydział Matematyki i Informatyki UJ
➡️ when? Thursday (16 April 2026) at 14.15
➡️ Talk: “Making Extremely High-Dimensional Matrix Computations Scalable”

Abstract:
Modern machine learning increasingly operates in extremely high-dimensional regimes, where even basic matrix operations become computationally prohibitive due to their quadratic memory and time complexity. We show that for Laplace kernel matrices such trade-offs are not necessary. We introduce a structured representation that enables exact computation without materializing them explicitly with O(n) memory and time complexity. Conceptually, our approach provides an alternative to both low-rank and randomized methods: it retains the expressivity of full-rank operators while achieving scalability through structure rather than approximation.

Witness it in action! 🔥

08/04/2026

🎤 Ready for fresh ideas? This week’s GMUM seminar features a talk you won’t want to miss.

➡️ who? Jan Miksa
➡️ where? Wydział Matematyki i Informatyki UJ
➡️ when? Thursday (9 April 2026) at 14.15
➡️ Talk: “BARRIER: Bounded Activation Regions for Robust Information Erasure”

Abstract:
Selectively erasing unwanted concepts, such as unsafe or copyrighted content, is increasingly critical for large computer vision models. Existing unlearning methods manipulate static weights, but dense concept entanglement in standard weight spaces frequently causes collateral damage to retained knowledge. To prevent this, we propose BARRIER (Bounded Activation Regions for Robust Information Erasure), a novel framework that shifts the intervention locus from static weights to the dynamic latent geometry of hidden layer activations. Using Singular Value Decomposition (SVD) and Interval Arithmetic, BARRIER projects activations into a low-rank subspace to establish stable constraints for retained concepts. This provides theoretical guarantees that explicitly bound functional drift, preserving essential knowledge while enabling the surgical suppression of unwanted concepts. As architecture-agnostic, BARRIER requires no additional trainable parameters. Evaluations demonstrate state-of-the-art efficacy across diverse paradigms. In classification, it successfully suppresses targeted classes while maintaining robust test performance. In diffusion models, BARRIER eradicates specific visual classes and explicit content without degrading generative quality significantly, matching or outperforming weight-centric baselines.

See it live in action! 🔥