About me
Hello! I am an AI Researcher at the Flowers AI and CogSci Lab (INRIA, Bordeaux) led by Pierre-Yves Oudeyer. I studied how psychology and cultural evolution can be adequately leveraged to better characterize, evaluate, and build LLMs and RL agents.
My last research project focused on Synthetic Data and Model Collapse, where we used cultural evolution methodologies to identify the specific training data properties that mitigate or foster degradation in recursive training loops. A great part of my research studied how LLMs encode and express culture and values under trivial context changes (fuzzing). This includes an influential position paper, and the StickToYourRole Leaderboard, which uses psychological theories and methodology to evaluate the stability of value expression in LLM-simulated populations. You can see my thesis here.
Previously, I worked in Microblink on developing DL models, primarily for vision tasks such a OCR, and then as an engineer in the Flower Team on levering learning progress for autonomous goal-based exploration of RL-agent in the visual domain.
In my spare time I like to play with old vintage bikes, such as my Gitane Champion du Monde from 1975. :)
Projects
You can see our recent research projects below:
- Recursive Training Loops in LLMs - How training data properties modulate degradation and change in synthetic LLM-generated data following recursive fine-tuning
- Stick to your Role! Stability of Personal Values Expressed in Large Language Models - a study of Personal Value stability expressed by LLM-simulated personas under trivial context changes (fuzzing)
- Stick to your Role! Leaderboard - a leaderboard based on and extending the StickToYourRole paper
- Large Language Models as superpositions of cultural perspectives - a positioning regarding the influence of context change on the expression of culture in LLMs
- The SocialAI school - a perspective based on developmental psychology, and a tool facilitating research into socio-cognitive abilities of RL- and LLM- based agents
- GRIMGEP: Learning Progress for Robust Goal Sampling in Visual Deep Reinforcement Learning - study of intrinsically motivated goal-based exploration with RL
Other projects:
- GrgoBot - an Agentic RAG agent that explains my research
- When LLMs Play the Telephone Game - Cumulative Changes and Attractors in Iterated Cultural Transmissions between LLMs
- Autotelic LLM-based exploration for goal-conditioned RL - open-ended exploration with LLMs and RL
Publications
See my Google ScholarPeer-reviewed:
- Grgur Kovač*, Jérémy Perez*, Rémy Portelas, Peter Ford Dominey, and Pierre-Yves Oudeyer. 2025. Recursive Training Loops in LLMs: How training data properties modulate distribution shift in generated data? In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP 2025, Oral)
- Jérémy Perez, Grgur Kovač, Corentin Léger, Cédric Colas, Gaia Molinaro, Maxime Derex, Pierre-Yves Oudeyer, and Clément Moulin-Frier (2025). ‘When LLMs Play the Telephone Game: Cultural Attractors as Conceptual Tools to Evaluate LLMs in Multi-turn Settings’. In: The Thirteenth International Conference on Learning Representations (ICLR 2025)
- Grgur Kovač, Rémy Portelas, Masataka Sawayama, Peter Ford Dominey, and Pierre-Yves Oudeyer (2024b). ‘Stick to your Role! Stability of Personal Values Expressed in Large Language Models’. In: Proceedings of the Annual Meeting of the Cognitive Science Society. Vol. 46 (CogSci 2024)
- Grgur Kovač, Rémy Portelas, Masataka Sawayama, Peter Ford Dominey, and Pierre-Yves Oudeyer (Aug. 2024a). ‘Stick to your role! Stability of personal values expressed in large language models’. In: PLOS ONE 19.8
- Grgur Kovač, Rémy Portelas, Peter Ford Dominey, and Pierre-Yves Oudeyer (2024). ‘The SocialAI school: a framework leveraging developmental psychology toward artificial socio-cultural agents’. In: Frontiers in Neurorobotics Volume 18 - 2024
- Grgur Kovač, Adrien Laversanne-Finot, and Pierre-Yves Oudeyer (2022). ‘Grimgep: learning progress for robust goal sampling in visual deep reinforcement learning’. In: IEEE Transactions on Cognitive and Developmental Systems 15.3
Preprints:
- Grgur Kovač, Masataka Sawayama, Rémy Portelas, Cédric Colas, Peter Ford Dominey, and Pierre-Yves Oudeyer (2023). ‘Large language models as superpositions of cultural perspectives’. In: arXiv preprint arXiv:2307.07870
Workshops:
- Grgur Kovač*, Rémy Portelas*, Katja Hofmann, and Pierre-Yves Oudeyer (June 2021). ‘SocialAI 0.1: Towards a Benchmark to Stimulate Research on Socio-Cognitive Abilities in Deep Reinforcement Learning Agents’. In: NAACL. Accepted at NAACL ViGIL Workshop 2021. Mexico City, Mexico (Spotlight)
- Grgur Kovač, Rémy Portelas, Peter Ford Dominey, and Pierre-Yves Oudeyer (July 2023). ‘The SocialAI School: Insights from Developmental Psychology Towards Artificial Socio-Cultural Agents’. In: TOM 2023 -First Workshop on Theory of Mind in Communicating Agents - ICML 2023 Workshop. Honolulu (Hawaii), United States
- Guillaume Pourcel, Thomas Carta, Grgur Kovač, and Pierre-Yves Oudeyer (2024). ‘Autotelic LLM-based exploration for goal-conditioned RL’. In: Intrinsically Motivated Open-ended Learning Workshop at NeurIPS 2024