Theory, Practice, and Security of Foundation Models (Boninchi)

29 Sep 2025, 08:05 → 3 Oct 2025, 20:05 Europe/Zurich

Villa Boninchi

Foundation models have emerged as a transformative technology in AI, revolutionizing tasks across vision, language, and multimodal domains. These models, trained on massive datasets and capable of self-supervised learning, serve as a backbone for a wide array of applications, from content generation to complex decision-making systems. Despite remarkable practical results, many fundamental theoretical questions remain unanswered. There is an ongoing debate among researchers regarding the true potential and limitations of these models, with some claiming they approach general artificial intelligence (AGI) capabilities, while others remain skeptical due to the lack of robust theoretical understanding.

This workshop seeks to bring together leading experts from academia and industry to foster a deeper understanding of the theoretical foundations of foundation models while simultaneously addressing practical advances and real-world deployments. A particular emphasis will be also placed on understanding and improving the security and robustness of these models. Topics such as adversarial vulnerabilities, watermarking for copyright protection, and robustness to distributional shifts will be core aspects of the discussion. Given the increasing reliance on foundation models in critical applications, this workshop will also explore methods to secure these models against malicious attacks.

This workshop will feature interactive sessions with invited talks by world-class experts in theory, practice, and security of foundation models. It will provide a unique platform for researchers and practitioners to engage in discussions that shape the next generation of robust, secure, and theoretically sound foundation models.

Bus_station_to_Venue.png

Hotel to Workshop (1).pdf

Monday, 29 September

09:00 → 09:30 Welcome & intro

Speaker: Slava Voloshynovskiy (UniGE)

09:30 → 10:40 Large Models in Low Precision: Power, Limits and Scaling Laws

Speaker: Dan Alistarh ( ISTA, Austria)

10:40 → 11:00 Coffee break ☕️🫖🥐

11:00 → 12:10 Short-term attention enables long term memorization & Stochastic activations

Speaker: Hervé Jégou (META)

12:10 → 13:10 Lunch

13:10 → 14:10 AI through the Information-Theoretic lens

Speaker: Slava Voloshynovskiy (UniGE)

14:10 → 14:30 Coffee break☕️🫖

14:30 → 15:00 TwinTURBO: Semi-Supervised Fine-Tuning of Foundation Models via Mutual Information Decompositions

Speaker: Quetant Guillaume (UNIGE)

15:00 → 15:30 Llama4 in 4 bits

Speaker: Brian Pulfer (University of Geneva)

15:30 → 15:50 Coffee break☕️🫖

15:50 → 17:00 Diffusion LM and Compression

Speaker: Pavlo Molchanov (NVIDIA, USA)

17:00 → 18:00 Apero on Lake Board: discussions/networking/free time

18:00 → 18:30 Transfer to Social Event

18:30 → 22:00 Dinner

Leave for https://chateaudif.ch: [Google Maps]https://maps.app.goo.gl/HpDQCu4QvUxoy5dT8

Tuesday, 30 September

09:00 → 09:30 ☕️ 🫖 🥐

09:30 → 10:40 The Watermark Awakens

Speaker: Pierre Fernandez (Meta, France)

10:40 → 11:00 ☕️ 🫖

11:00 → 12:10 The power of small language models

Speaker: Pavlo Molchanov (NVIDIA, USA)

12:10 → 13:40 Lunch

13:40 → 14:10 Handcrafted vs. Learned: A Comparative Evaluation of Digital Watermarking Techniques

Speaker: Vitaliy Kinakh

14:10 → 14:40 Attacking and Defending Vision Foundation Models

Speaker: Yury Belousov

14:40 → 15:00 ☕️ 🫖

15:00 → 15:30 TBD

Speaker: Brian Pulfer (University of Geneva)

15:30 → 15:50 ☕️ 🫖

15:50 → 17:00 TBD

Speaker: Seyed Mohsen Moosavi-Dezfooli (Apple, CH)

17:00 → 17:20 ☕️ 🫖

17:20 → 18:00 Round table - industrial perspective of model security and privacy

Wednesday, 1 October

10:45 → 12:25 Boat trip to Yvoire

https://maps.app.goo.gl/5xWLTmPKiUcsABdm6

12:30 → 16:30 Lunch & Discussion

https://hdpy.fr/fr/

16:45 → 18:20 Boat trip -Return to Geneva

Thursday, 2 October

09:00 → 09:30 ☕️🫖🥐

09:30 → 10:40 Generative World Models for Sample-Efficient Reinforcement Learning

Speakers: Eloi Alonso, Vincent Micheli

10:40 → 11:00 ☕️🫖

11:00 → 12:10 Energy-Based Transformers and the Future of Scaling

Speaker: Alexi Gladstone (University of Illinois Urbana-Champaign/META, the USA)

12:10 → 13:40 Lunch

13:40 → 14:30 Mini-Bang Detection and Interpretation: Benchmarking Reasoning on Emergent and Open-Ended Dynamics

Speaker: Andrey Ustuzanin (Constructor University, Germany)

14:30 → 14:50 ☕️🫖

14:50 → 15:40 Improving mid-resource language performance in open LLMs

Speaker: Hanna Yukhymenko (Hugging Face, CH)

15:40 → 16:00 ☕️🫖

16:00 → 16:30 TBD

Speaker: Yury Belousov

16:30 → 17:00 Binary Diffusion Probabilistic Model

Speaker: Vitaliy Kinakh

17:00 → 18:00 Discussion & Brainstorm & Free time

Friday, 3 October

09:00 → 09:30 ☕️🫖🥐

09:30 → 10:30 DINOv3

Speaker: Federico Baldassarre (Meta, France)

10:30 → 11:30 How perception got solved (albeit unsatisfyingly)

Speakers: Lucas Beyer, Xiaohua Zhaị

11:30 → 11:45 ☕️🫖

11:45 → 12:45 Reasoning by thinking compositionally

Speaker: Emmanuel Abbe (EPFL / Apple, CH)

12:45 → 14:00 Lunch

14:00 → 15:00 World Models for Code

Speaker: Taco Cohen (Meta, the UK)

15:00 → 15:20 ☕️🫖

15:20 → 16:20 The free transformer

Speaker: François Fleuret (University of Geneva /Meta, CH/France)

16:20 → 16:40 ☕️🫖

16:40 → 17:10 Vision-language models: in-context learning and adaptation for radio-astronomy

Speaker: Mariia Drozdova (Universite de Geneve (CH))

17:10 → 17:40 Radio Astronomy in the Era of Large Models

Speaker: Erica Lastufka (University of Geneva, CH)

17:40 → 18:00 Summary/Conclusion/Outlook

Speaker: Slava Voloshynovskiy (UniGE)

18:05 → 20:00 Afterparty by subscription