Paper: Rethinking RL for LLM Reasoning via Sparse Policy Selection

mark_k · x · 2026-08-16

Paper Link: https://t.co/FGelGZFpXR

The study 'Rethinking RL for LLM Reasoning' confirms that RL improves reasoning not by learning new strategies but via sparse policy selection. Token-level analysis reveals RL's benefits are concentrated at just 1–3% high-entropy token positions, where promoted tokens are always within the base model's Top-5.

Based on this, the authors propose ReasonMaxxer, an embarrassingly cheap post-training method. It uses the base model's own entropy to identify decision points and applies contrastive loss only there, requiring no online generation. Experiments show ReasonMaxxer matches or exceeds full RL performance across 3 model families, 6 scales, and 6 math benchmarks.

Related event: Paper: RL Teaches LLMs No New Strategies, Gains Reproduced at 1/1000th Compute(4 posts)→

Original post →

More from Research

Research channel →