SHAPE improves average math reasoning accuracy by 3% while cutting token use by 30% through stage-aware hierarchical advantage and entropy-driven token redistribution.
The KL divergence penalty coefficient is set to 0 to prioritize direct reward optimization, relying on the clipping mechanism for policy constraints
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
SHAPE: Stage-aware Hierarchical Advantage via Potential Estimation for LLM Reasoning
SHAPE improves average math reasoning accuracy by 3% while cutting token use by 30% through stage-aware hierarchical advantage and entropy-driven token redistribution.