Real-Time Aligned Reward Model beyond Semantics

In three linesR2M (Real-Time Aligned Reward Model) introduces a lightweight RLHF framework to mitigate reward overoptimization. Instead of relying solely on semantic representations, R2M leverages evolving hidden states from the policy model to align with real-time distribution shifts during reinforcement learning training.

Summary generated by Claude — human-verified