AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

GeoAlign: Geometric Rollout Curation for Robust LLM Reinforcement Learning

Ting Zhou, Zhenqing Ling, Zhao Yiyang, Ying Shen, Daoyuan Chen

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 GeoAlign rollout curationBase-GRPO Qwen3-1.7B · DAPO-Mathaverage accuracy: 39.81 -> 40.44Improved
add GeoAlign rollout curationBase-GRPO Qwen3-4B · DAPO-Mathaverage accuracy: 54.78 -> 55.94Improved
add GeoAlign rollout curationBase-GRPO Qwen3-1.7B · HH-RLHFaverage score: 0.8354 -> 0.8885Improved
add GeoAlign rollout curationBase-GRPO Qwen3-4B · HH-RLHFaverage score: 0.8672 -> 0.8894Improved
add GeoAlign under 5% reward flip noiseBase-GRPO Qwen3-1.7B · DAPO-Mathaverage accuracy: 36.98 -> 38.20Improved
add GeoAlign under 10% reward corruptionBase-GRPO Qwen3-1.7B · HH-RLHF proxy rewardproxy reward: 0.0808 -> 0.0820Improved
add Anomaly-Amplify (oversample outliers) instead of GeoAlign curation (ablation)GeoAlign · DAPO-Mathaverage accuracy: 40.44 -> 39.71Hurt
add Anomaly-Amplify (oversample outliers) instead of GeoAlign curation (ablation)GeoAlign · HH-RLHF proxy rewardproxy reward: 0.0848 -> 0.0834Hurt
use cumulative rather than normalized anomaly score aggregation (ablation)GeoAlign · DAPO-Mathaverage accuracy: 39.48 -> 40.44Improved
use cumulative rather than normalized anomaly score aggregation (ablation)GeoAlign · HH-RLHF proxy rewardproxy reward: 0.0837 -> 0.0848Improved
vary projector depth from 2 to 5 layers (ablation)GeoAlign projector · HH-RLHF validation accuracy (Qwen3-4B)accuracy: 87.02% / 86.98% / 87.09% / 86.40%No clear effect
increase density threshold alpha beyond task-specific peak (0.15 vs 0.12) (ablation)GeoAlign · HH-RLHF proxy rewardproxy reward: 0.0848 -> 0.0839Hurt

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