DAPO: An Open-Source LLM Reinforcement Learning System at Scale

DAPO turns four failure diagnoses in long-chain-of-thought reinforcement learning into four concrete interventions: asymmetric clipping, informative-prompt sampling, token-level loss normalization, and gradual overlength penalties. Follow each control, then audit the cumulative AIME experiment and the compute claims it cannot settle.

Reading focus: Why a symmetric probability-ratio clip can stop rare, positively rewarded tokens from becoming much more likely. How dynamic sampling removes prompt groups whose identical rewards produce zero normalized advantage. What the cumulative 30-to-50 AIME result establishes, and why it does not isolate four independent causal effects.

NeurIPS 2025. Qiying Yu et al.. 13 min read, hard difficulty.