Research map / Optimization / Optimizers
Clipping Bottleneck: Stabilizing RLVR via Stochastic Recovery of Near-Boundary Signals
Shuo Yang, Jinda Lu, Chiyu Ma, Kexin Huang, Haoming Meng, Qihui Zhang, Yuyang Liu, Bolin Ding, Guoyin Wang, Li Yuan, Jingren Zhou
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 NSR to DAPO | DAPO · AIME24 (Qwen2.5-Math-7B) | Pass@1: 35.83 -> 40.76 | Improved |
| add NSR to DAPO | DAPO · AIME24 (Qwen3-8B-Base) | Pass@1: 37.29 -> 43.65 | Improved |
| add NSR to GSPO | GSPO · AIME24 (Qwen3-30B-A3B-Base) | Pass@1: 54.17 -> 58.65 | Improved |
| add NSR to DAPO | DAPO · AMC (Qwen2.5-Math-7B) | Pass@1: 69.28 -> 76.74 | Improved |
| add NSR to DAPO | DAPO · GPQA (Qwen2.5-Math-7B) | accuracy: 39.27 -> 42.99 | Improved |
| add NSR to GSPO | GSPO · GPQA (Qwen3-30B-A3B-Base) | accuracy: 58.96 -> 59.22 | No clear effect |
| add NSR to GSPO | GSPO · MMLU-Pro (Qwen3-30B-A3B-Base) | accuracy: 72.08 -> 74.27 | Improved |
| replace stochastic rescue with binary admission (ablation) | NSR · AIME24 (Qwen2.5-Math-7B) | Pass@1: 40.76 -> 37.84 | Hurt |
| replace stochastic rescue with deterministic decay k=2 (ablation) | NSR · AIME24 (Qwen2.5-Math-7B) | Pass@1: 40.76 -> 38.68 | Hurt |
| use deterministic decay k=4 instead of NSR (ablation) | NSR · AIME24 (Qwen2.5-Math-7B) | Pass@16: 57.70 -> 54.08 | Hurt |
| add binary admission of out-of-bound tokens to DAPO (ablation) | DAPO · AIME24 (Qwen2.5-Math-7B) | Pass@1: 35.83 -> 37.84 | Improved |
| remove advantage normalization from DAPO (ablation) | DAPO · AIME24 | peak Pass@1: 37% -> 38% | No clear effect |
Closest papers by meaning
- What are Key Factors for Updates in RL for LLM Reasoning? (NeurIPS 2026 · similarity 0.89)
- You Only Need Minimal RLVR Training: Extrapolating LLMs via Rank-1 Trajectories (NeurIPS 2026 · similarity 0.88)
- The Reasoning Boundary Paradox: How Reinforcement Learning Constrains Language Models (NeurIPS 2026 · similarity 0.88)
- Not Only Where, But When: Temporal Scheduling for RLVR (NeurIPS 2026 · similarity 0.87)
- On the Direction of RLVR Updates for LLM Reasoning: Identification and Exploitation (ICLR 2026 · similarity 0.87)
- Detecting RLVR Training Data via Structural Convergence of Reasoning (NeurIPS 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.
