Circuit-Aware Reward Training: A Mechanistic Framework for Longtail Robustness in RLHF

Open in new window