Research map / Reinforcement learning / RL from feedback
FAPO: Flawed-Aware Policy Optimization for Efficient and Reliable Reasoning
Yuyang Ding, Chi Zhang, Juntao Li, Haibin Lin, Xin Liu, Min Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add flawed-positive penalty (FAPO) to GRPO baseline | Qwen2.5-32B RL · AIME24 | accuracy: 38.9 -> 42.4 | Improved |
| add flawed-positive penalty (FAPO) to GRPO baseline | Qwen2.5-32B RL · LiveCodeBench | accuracy: 28.6 -> 33.6 | Improved |
| add flawed-positive penalty (FAPO) to GRPO baseline | Qwen2.5-32B RL · Avg over 6 benchmarks | average accuracy: 51.0 -> 54.8 | Improved |
| add flawed-positive penalty (FAPO) to GRPO baseline | Qwen2.5-32B RL · AIME24 flawed-positive ratio (LLM judge) | flawed-positive ratio (%): 15.5 -> 7.1 | Improved |
| penalize flawed positives using Qwen3-32B detector in preliminary RL (ablation) | Qwen2.5-Math-7B RLVR · AIME24 | accuracy | Improved |
| add process reward (step-distance penalty) to GenRM outcome reward | FAPO-GenRM-4B · FlawedPositiveBench | F1: 81.2 -> 89.4 | Improved |
| RL-train GenRM from Qwen3-4B-Instruct | Qwen3-4B-Instruct · ProcessBench | average F1: 77.1 -> 83.3 | Improved |
| use a stronger GenRM (FAPO-GenRM vs base Qwen3-4B-Instruct) inside RL (ablation) | FAPO-Reasoning · AIME24/AIME25/GPQA-Diamond | accuracy during RL | Improved |
| use step-ratio process reward (proportion of correct steps before first error) instead of FAPO penalty (ablation) | FAPO-Reasoning · AIME24/AIME25/GPQA-Diamond | accuracy during RL | Hurt |
| increase lambda penalty from 0 to 1/3 (ablation) | FAPO-7B · 7B benchmark average | performance: 32.1 -> 34.6 | Improved |
| use default lambda=1 instead of 1/3 (ablation) | FAPO-7B · 7B benchmark average | performance: 34.6 -> 36.8 | Improved |
| use more aggressive lambda=-1/3 (shift earlier) (ablation) | FAPO-7B · 7B benchmark average | performance: 36.8 -> 39.6 | Improved |
Closest papers by meaning
- DAPO : Improving Multi-Step Reasoning Abilities of Large Language Models with Direct Advantage-Based Policy Optimization (NeurIPS 2025 · similarity 0.91)
- Decoupling Reasoning and Confidence: Resurrecting Calibration in Reinforcement Learning from Verifiable Rewards (ICML 2026 · similarity 0.91)
- Resource-Efficient Reinforcement for Reasoning Large Language Models via Dynamic One-Shot Policy Refinement (ICML 2026 · similarity 0.90)
- Advancing LLM Reasoning with Natural Language and Numerical Feedback (ICML 2026 · similarity 0.90)
- Your Language Model is Its Own Critic: Reinforcement Learning with Value Estimation from Actor’s Internal States (NeurIPS 2026 · similarity 0.90)
- Optimizing Anytime Reasoning via Budget Relative Policy Optimization (NeurIPS 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
