AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Advantage Collapse in Group Relative Policy Optimization: Diagnosis and Mitigation

Xixiang He, Qiyao Sun, Ao Cheng, Xingming Li, Xuanyu Ji, Hailun Lu, Runke Huang, Qingyong Hu

ICML 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants · 0 citations (OpenAlex)

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 AVSPO virtual samples to GRPOGRPO, Qwen2.5-0.5B to 14B · six math benchmarksadvantage collapse rate: 0.28-0.45 -> 0.11-0.18Improved
add AVSPO to GRPOQwen2.5-Math-7B · MATH-500pass@1 accuracy: 65.0 -> 74.1Improved
add AVSPO to GRPO on out-of-domain MMLU-ProQwen2.5-3B · MMLU-Proaccuracy: 37.4 -> 38.6No clear effect
use stratified virtual rewards instead of random uniform (ablation)AVSPO, Qwen2.5-Math-1.5B · MATH-500accuracy: 62.1 -> 67.2Improved
use stratified virtual rewards instead of fixed partial credit (ablation)AVSPO, Qwen2.5-Math-1.5B · MATH-500accuracy: 63.5 -> 67.2Improved
augment all-incorrect groups in addition to all-correct groups (ablation)AVSPO, Qwen2.5-Math-1.5B · MATH-500accuracy: 60.8 -> 67.2Improved
augment all-correct groups in addition to all-incorrect groups (ablation)AVSPO, Qwen2.5-Math-1.5B · MATH-500accuracy: 63.2 -> 67.2Improved
use adaptive threshold instead of fixed tau=0.5 (ablation)AVSPO, Qwen2.5-Math-1.5B · MATH-500final accuracy: 65.4 -> 67.2Improved
increase alpha from 0.5 to 1.0 (ablation)AVSPO · MATH-500accuracy: drops 2.1%Hurt
decrease alpha from 0.5 to 0.3 (ablation)AVSPO · MATH-500ACR: residual ACR = 0.24Hurt
vary initial threshold in [0.3,0.7] and eta in [0.005,0.02] (ablation)AVSPO · MATH-500accuracy: less than 1% accuracy variationNo clear effect
augment collapsed groups (AVSPO) instead of filtering them (Filter-Drop) (ablation)GRPO, Qwen2.5-Math-7B · MATHaccuracy: 67.1 -> 74.1Improved

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