Preventing Entropy Collapse: SAF Framework Enhances LLM Math Reasoning

Yifan Ding · hf · 2026-08-03

This paper introduces the Stable Advantage Fusion (SAF) framework to address challenges in combining reinforcement learning with on-policy distillation (OPD). The authors find that fusing RLVR (Reinforcement Learning with Verifiable Rewards) and OPD advantages with a fixed coefficient triggers entropy collapse due to magnitude and temporal miscalibrations.

SAF resolves these issues via a lightweight, four-stage pipeline applied to the OPD advantage: a sparsify-then-compress mechanism for magnitude control paired with a warm-up-then-anneal mechanism for temporal control. Evaluated on Qwen3-1.7B/4B/8B across seven math reasoning and code generation benchmarks, SAF consistently prevents entropy collapse, improving aggregate scores by 0.51-2.70% over fixed-coefficient fusion while ensuring stable training.

Original post →

More from Research

Research channel →