AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Group Verification-based Policy Optimization for Interactive Coding Agents

Silong Dai, Changzhi Sun, Haolun Wu, Huanran Zheng, Tao Ji, Junchi Yan, Yuanbin Wu, Dell Zhang, Xiaoling Wang, Xuelong Li

ICLR 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants

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
use sequence-mean-token-mean instead of token-mean aggregation (ablation)GVPO · AppWorld Test-CTGC: 34.6 -> 49.4Improved
use sequence-mean-token-mean instead of sequence-mean-token-sum aggregation (ablation)GVPO · AppWorld Test-CTGC: 42.8 -> 49.4Improved
use asymmetric clipping instead of symmetric clipping (ablation)GVPO · AppWorld Test-CTGC: 40.5 -> 49.4Improved
do not apply std normalization to advantages (ablation)GVPO · AppWorld Test-CTGC: 42.8 -> 49.4Improved
match micro-batch and group sizes (ablation)GVPO · AppWorld Test-CTGC: 43.1 -> 49.4Improved
use multiplicative shaping instead of additive-only shaping (ablation)GVPO · AppWorld Test-CTGC: 39.8 -> 49.4Improved
add multiplicative shaping on Test-N (ablation)GVPO · AppWorld Test-NTGC: 64.9 -> 72.6Improved
use asymmetric clipping on Test-N (ablation)GVPO · AppWorld Test-NTGC: 60.1 -> 72.6Improved
lower shaping coefficient b from 0.2 to 0.1 (ablation)GVPO · AppWorld Test-CTGC: 49.4 -> 53.1Mixed
raise shaping coefficient b from 0.2 to 0.4 (ablation)GVPO · AppWorld Test-NTGC: 72.6 -> 67.9Hurt
add process-verifiable shaping to outcome-only RL (GVPO vs LOOP)LOOP · AppWorld Test-CTGC: 45.7 -> 49.4Improved
RL fine-tune Qwen2.5-32B with GVPO versus zero-shot promptingQwen2.5-32B-Instruct · AppWorld Test-CTGC: 18.9 -> 49.4Improved

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