Séminaire ayant lieu dans le cadre du Doctorat en informatique cognitive, en collaboration avec le centre de recherche CRIA              

TITRE : Multimodal evaluation/State computation

Yoav ARTZI

Jeudi le 3 décembre 2026 à 10h30

Local PK-5115 (Il est possible d'y assister en virtuel en vous inscrivant ici)                 

RÉSUMÉ

This talk briefly describes three distinct projects. The first two focus on evaluating multi-modal capabilities in models through the evaluation of dual perspective reasoning and fine-grained manipulation. For the former, we focus on spatial reasoning across egocentric and allocentric observations, following setups designed for the evaluation of cognitive development in children. The latter utilizes knot manipulation, a scenario well modeled by mathematical theory, which provides a clear avenue to gauge complexity and generalization. The last part of the talk is focused on a fundamental capability critical for cross-modal reasoning: the ability to separate between how one represents state and one's prediction of outputs. We hypothesize that mechanically separating the two in LLMs is beneficial and show this in the context of LLM pre-training.

BIOGRAPHIE

Yoan ARTZI is Associate Professor in the Department of Computer Science and Cornell Tech at Cornell University and visiting research faculty at Google DeepMind. His research is on language modeling and learning in interactive and situated scenarios.

RÉFÉRENCES 

Chen, Z., & Artzi, Y. (2025). Knot so simple: A minimalistic environment for spatial reasoning. Advances in Neural Information Processing Systems: Datasets and Benchmarks Track (NeurIPS DB). arXiv 

Monea, G., Godey, N., Brantley, K., & Artzi, Y. (2026). The state-prediction separation hypothesis. arXiv. arXiv