Event
A Rationale-Guided Framework for Human Reward Modeling
1 reports · 1 sources
Anshul Nayak et al. propose a rationale-guided preference-based reinforcement learning (PbRL) framework that learns a human reward model from trajectory-level feedback and combines the learned reward with the environment reward for policy optimization. The framework introduces rationale labels in addition to preference labels, indicating the reasons why behaviors are preferred, so that the reward model captures both human choices and the grounds behind them. The paper states that, compared with reward learning using preferences alone, this method provides additional supervision and improves the interpretability of the policy.
Report timeline
- Reason-Guided Human Reward Modeling for Preference-Based Reinforcement Learning
This paper proposes a reason-guided preference-based reinforcement learning (PbRL) framework that learns a human reward model from trajectory-level feedback, introducing reason tags that explain why a behavior is preferred in addition to preference labels. Compared with preference-only reward learning, this method provides additional supervision for learning human-aligned rewards and improves the interpretability of the policy. The learned human reward is combined with the environment reward for policy optimization, enabling the agent to improve task performance while producing more interpretable behavior that better matches human preferences and aligns with human intent.
arXiv cs.RO ·