Rethinking Entropy Interventions in RLVR: An Entropy Change Perspective
This work addresses a critical risk in RLVR training for LLM reasoning, offering a novel solution to improve generalization, though it is incremental relative to existing entropy-intervention methods.
The paper tackles the problem of entropy collapse in Reinforcement Learning with Verifiable Rewards (RLVR) for LLM reasoning, introducing STEER, a method that stabilizes token-level entropy changes and achieves stronger downstream performance across mathematical reasoning benchmarks.
While Reinforcement Learning with Verifiable Rewards (RLVR) can enhance LLM reasoning, its training process poses a critical risk: entropy collapse. This phenomenon is a rapid loss of policy diversity, stemming from the exploration-exploitation imbalance and leading to a lack of generalization. Recent entropy-intervention methods aim to prevent \coloredtext{entropy collapse}, yet their underlying mechanisms remain unclear. In this paper, we conduct a quantitative analysis to reveal token-level entropy changes and how existing entropy intervention methods help avoid entropy collapse. Our findings point out a fundamental limitation of existing methods: they attempt to control entropy dynamics indirectly. By only affecting related factors, such as the advantage signal and generation probability, their effectiveness is inherently limited and could potentially fail. To address this limitation, we introduce an entropy-change-aware reweighting scheme, namely Stabilizing Token-level Entropy-changE via Reweighting (STEER), that adaptively stabilizes entropy dynamics through fine-grained token-level adjustments. Our approach mitigates over-exploitation while fostering robust exploration. Extensive experiments demonstrate that STEER significantly mitigates entropy collapse, stabilizes entropy dynamics, and achieves stronger downstream performance across various mathematical reasoning benchmarks \footnote{Our code is available at https://github.com/zz-haooo/STEER.