---
title: "DAPO: An Open-Source LLM Reinforcement Learning System at Scale"
description: "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."
canonical_url: "https://fanout.sh/daily/2026-10-26-dapo-llm-reinforcement-learning"
md_url: "https://fanout.sh/daily/2026-10-26-dapo-llm-reinforcement-learning.md"
last_updated: "2026-10-26"
access: "public"
---

# 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.

## Paper details

- Authors: Qiying Yu et al.

- Venue: NeurIPS 2025

- Track: AI Research

- Difficulty: Hard

- Reading time: 13 min

- Original paper: https://arxiv.org/abs/2503.14476

## What you will learn

- 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.

## Continue exploring

[Review group-relative advantages](https://fanout.sh/daily/2026-08-02-deepseekmath-grpo): Start with GRPO's critic-free group normalization before inspecting the four training controls DAPO adds around it.

---
This representation contains public Fanout content only. Protected Pro lessons, account data, billing, checkout, and pricing are not included.

Browse the public content map: https://fanout.sh/sitemap.md