Causally Robust Reward Learning from Reason-Augmented Preference Feedback

ABSTRACT

Reward learning from human preferences often suffers from spurious correlations, leading agents to develop brittle and misaligned behaviors. The authors present a framework that integrates causal inference with reason-augmented feedback, forcing the agent to learn the actual causal drivers of a preferred outcome rather than superficial patterns. Validated on complex robotic manipulation tasks, this method significantly improves the robustness of agentic systems in novel environments.