Interpreting Reinforcement Learning Agents with Susceptibilities

TL;DR

This paper introduces susceptibilities as a neural network interpretability technique for deep reinforcement learning, revealing internal model features beyond policy development, validated through a gridworld model and activation-steering.

Contribution

It generalizes susceptibilities to deep RL, demonstrating their utility in understanding internal model features and extending the framework to RLHF post-training.

Findings

Susceptibilities reveal internal features not detectable by policy analysis.

Validated susceptibilities with activation-steering in a gridworld model.

Discussed extension of susceptibilities to RLHF post-training.

Abstract

Susceptibilities are a technique for neural network interpretability that studies the response of posterior expectation values of observables to perturbations of the loss. We generalize this construction to the setting of the regret in deep reinforcement learning and investigate the utility of susceptibilities in a simple gridworld model that nevertheless exhibits non-trivial stagewise development. We argue that susceptibilities reveal internal features of the development of the model in parameter space that one cannot detect purely by studying the development of the learned policy. We validate these results with activation-steering, and discuss the framework's extension to RLHF post-training.