A Few Bad Apples Spoil the Bunch: Preventing Global Entropy Collapse Driven by a Small Set of Tokens in LLM Reasoning

摘要

Reinforcement Learning with Verifiable Rewards (RLVR) and Reinforcement Learning from Internal Feedback (RLIF) often fail to benefit from test-time compute due to entropy collapse and the resulting loss of reasoning diversity. We show that this collapse is driven not by uniform entropy decay, but by premature overconfidence at a small number of structurally critical decision points. Based on a token-level analysis of GRPO-style policy optimization, we propose SCOPE (Structural Collapse-aware Optimization via Partial Entropy control), which assigns each generated token a redistribution score and applies selective KL regularization to only the top \sim 5% of tokens under this score. Across model scales and architectures on math reasoning benchmarks, SCOPE consistently improves performance under both RLVR and RLIF settings, demonstrating that targeted entropy control at a vanishingly small subset of tokens is sufficient to sustain reasoning diversity and effective test-time scaling.