Publications

We present TDHook, an open-source, lightweight, generic interpretability framework based on tensordict and applicable to any torch model. It focuses on handling complex composed models and features ready-to-use methods for attribution, probing and a flexible get-set API for interventions.

By addressing Multi-Agent Deep Reinforcement Learning interpretability, we propose directions aiming to advance active topics such as team identification, swarm coordination and sample efficiency. This paper advocates for direct interpretability, generating post hoc explanations directly from trained models, offering scalable insights into agents' behaviour, emergent phenomena, and biases without altering models' architectures.

We propose contrastive sparse autoencoders (CSAE), a novel framework for studying pairs of game trajectories. Using CSAE, we are able to extract and interpret concepts that are meaningful to the chess-agent plans. We primarily focused on a qualitative analysis of the CSAE features before proposing an automated feature taxonomy. Furthermore, to evaluate the quality of our trained CSAE, we devise sanity checks to wave spurious correlations in our results.