sparse-critical-rlvr-token-analysis

【数据与研究】Analyzes how Reinforcement Learning from Verification Rewards (RLVR) improves reasoning by examining token-level probability distributions. Finds that >83% of token positions exhibit near-zero divergence—RL operates through sparse, targeted refinements. Cross-sampling experiments show 1.5-7.8% RL-selected tokens recover full gains, while reverting 5-…

独立 GitHub 快照 · 自动分类 · 以原作者文档为准