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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add GeoAlign rollout curation | Base-GRPO Qwen3-1.7B · DAPO-Math | average accuracy: 39.81 -> 40.44 | Improved |
| add GeoAlign rollout curation | Base-GRPO Qwen3-4B · DAPO-Math | average accuracy: 54.78 -> 55.94 | Improved |
| add GeoAlign rollout curation | Base-GRPO Qwen3-1.7B · HH-RLHF | average score: 0.8354 -> 0.8885 | Improved |
| add GeoAlign rollout curation | Base-GRPO Qwen3-4B · HH-RLHF | average score: 0.8672 -> 0.8894 | Improved |
| add GeoAlign under 5% reward flip noise | Base-GRPO Qwen3-1.7B · DAPO-Math | average accuracy: 36.98 -> 38.20 | Improved |
| add GeoAlign under 10% reward corruption | Base-GRPO Qwen3-1.7B · HH-RLHF proxy reward | proxy reward: 0.0808 -> 0.0820 | Improved |
| add Anomaly-Amplify (oversample outliers) instead of GeoAlign curation (ablation) | GeoAlign · DAPO-Math | average accuracy: 40.44 -> 39.71 | Hurt |
| add Anomaly-Amplify (oversample outliers) instead of GeoAlign curation (ablation) | GeoAlign · HH-RLHF proxy reward | proxy reward: 0.0848 -> 0.0834 | Hurt |
| use cumulative rather than normalized anomaly score aggregation (ablation) | GeoAlign · DAPO-Math | average accuracy: 39.48 -> 40.44 | Improved |
| use cumulative rather than normalized anomaly score aggregation (ablation) | GeoAlign · HH-RLHF proxy reward | proxy reward: 0.0837 -> 0.0848 | Improved |
| 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 reward | proxy reward: 0.0848 -> 0.0839 | Hurt |
Closest papers by meaning
- ROLLVERIFY: BRIDGING EFFICIENCY AND ACCURACY IN LONG-TAIL ROLLOUT REINFORCEMENT LEARNING (NeurIPS 2026 · similarity 0.88)
- SR-GRPO: Stable Rank as an Intrinsic Geometric Reward for Large Language Model Alignment (NeurIPS 2026 · similarity 0.87)
- $f$-GRPO & Beyond: Divergence-Based Reinforcement Learning Algorithms for General LLM Alignment (NeurIPS 2026 · similarity 0.87)
- Rethinking the Sampling Criteria in Reinforcement Learning for LLM Reasoning: A Competence-Difficulty Alignment Perspective (AAAI 2026 · similarity 0.87)
- Enhancing the Outcome Reward-based RL Training of MLLMs with Self-Consistency Sampling (NeurIPS 2025 · similarity 0.87)
- One-Way Policy Optimization for Self-Evolving LLMs (ICML 2026 · similarity 0.87)
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.
