H$^2$R: Hierarchical Hint-Based Reward for Mathematical Reasoning via Reinforcement Learning
Sub-question hints give RLVR denser reward for math reasoning.
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