AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

GPG: A Simple and Strong Reinforcement Learning Baseline for Model Reasoning

Xiangxiang Chu, Hailang Huang, Xiao Zhang, Fei Wei, Yong Wang

ICLR 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants · 125 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
replace GRPO loss with plain policy gradient and group-mean reward (std normalization), no KL/reference (ablation)GPG · Math avg (5 benchmarks, ablation on MATH-lighteval training)average accuracy: 43.7 -> 45.3Improved
remove std reward normalization (Fnorm=1) from GPG (ablation)GPG · Math avg (5 benchmarks, ablation on MATH-lighteval training)average accuracy: 45.3 -> 43.9Hurt
add accurate gradient estimation (alpha=B/(B-M)) to Fnorm=1 GPG (ablation)GPG · Math avg (5 benchmarks, ablation on MATH-lighteval training)average accuracy: 43.9 -> 47.8Improved
add valid-sample threshold beta_th=0.6 with resampling (ablation)GPG · Math avg (5 benchmarks, ablation on MATH-lighteval training)average accuracy: 47.8 -> 48.3Improved
add alpha correction to std-normalized GPG (ablation)GPG · Math avg (5 benchmarks, ablation on MATH-lighteval training)average accuracy: 45.3 -> 44.1Hurt
add a KL distribution constraint to GPG (ablation)GPG · Math reasoningaccuracyHurt
increase group size from 2 to 16 (ablation)GPG · Math reasoningaverage accuracyImproved
use GPG instead of GRPO on 7B baseQwen2.5-Math-7B · Math avg (5 benchmarks, ablation on MATH-lighteval training)average accuracy: 43.7 -> 48.3Improved
use GPG instead of DAPOQwen2.5-Math-7B · Math avg, DAPO comparisonaverage accuracy: 56.0 -> 57.7Improved
use GPG instead of GRPO for geometry reasoningQwen2.5-VL-3B-Instruct · GEOQA testaccuracy: 47.48 -> 51.33Improved
use GPG instead of GRPO on CV-BenchQwen2-VL-2B · CV-Benchtotal accuracy: 59.47 -> 76.15Improved
use GPG instead of GRPO for LISA groundingQwen2-VL-2B · LISAmIoU: 37.6 -> 51.8Improved

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