Action-Value Calibration via Continuation Sharing for Long-Horizon Agentic Reinforcement Learning
Shares action continuations to calibrate action values, giving fairer credit in long-horizon agent RL.
Explainer · on the wire
The explainer for this story is still on the wire.
The claim, the mechanism, and the figures land here once generated.
Source · arXiv