AI Paper Atlas

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

ICLR 2026 · Reinforcement learning › RL from feedback · approach: Scaling RL with verifiable rewards

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 changedWhereMeasuredEffect
add flawed-positive penalty (FAPO) to GRPO baselineQwen2.5-32B RL · AIME24accuracy: 38.9 -> 42.4Improved
add flawed-positive penalty (FAPO) to GRPO baselineQwen2.5-32B RL · LiveCodeBenchaccuracy: 28.6 -> 33.6Improved
add flawed-positive penalty (FAPO) to GRPO baselineQwen2.5-32B RL · Avg over 6 benchmarksaverage accuracy: 51.0 -> 54.8Improved
add flawed-positive penalty (FAPO) to GRPO baselineQwen2.5-32B RL · AIME24 flawed-positive ratio (LLM judge)flawed-positive ratio (%): 15.5 -> 7.1Improved
penalize flawed positives using Qwen3-32B detector in preliminary RL (ablation)Qwen2.5-Math-7B RLVR · AIME24accuracyImproved
add process reward (step-distance penalty) to GenRM outcome rewardFAPO-GenRM-4B · FlawedPositiveBenchF1: 81.2 -> 89.4Improved
RL-train GenRM from Qwen3-4B-InstructQwen3-4B-Instruct · ProcessBenchaverage F1: 77.1 -> 83.3Improved
use a stronger GenRM (FAPO-GenRM vs base Qwen3-4B-Instruct) inside RL (ablation)FAPO-Reasoning · AIME24/AIME25/GPQA-Diamondaccuracy during RLImproved
use step-ratio process reward (proportion of correct steps before first error) instead of FAPO penalty (ablation)FAPO-Reasoning · AIME24/AIME25/GPQA-Diamondaccuracy during RLHurt
increase lambda penalty from 0 to 1/3 (ablation)FAPO-7B · 7B benchmark averageperformance: 32.1 -> 34.6Improved
use default lambda=1 instead of 1/3 (ablation)FAPO-7B · 7B benchmark averageperformance: 34.6 -> 36.8Improved
use more aggressive lambda=-1/3 (shift earlier) (ablation)FAPO-7B · 7B benchmark averageperformance: 36.8 -> 39.6Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy