Research map / Label-efficient and robust learning / Recognition and long-tail
Single-Rollout Hidden-State Dynamics for Training-Free RLVR Data Selection
Jianghao Wu, Jianfei Cai, Weiqiang Wang, Jin Ye, Daniel F. Schmidt, Yasmeen George
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 quality weighting and RIRS-based coverage (SHIFT) instead of random selection, 2% budget | RLVR data selection (Qwen2.5-Math-1.5B) · MATH-500 / AMC | Pass@1 accuracy: 53.73 -> 62.67 (MATH-500); 25.78 -> 38.55 (AMC) | Improved |
| select with SHIFT instead of CoreSet in question-embedding space, 2% budget | RLVR data selection (Qwen2.5-Math-1.5B) · MATH-500 | Pass@1 accuracy: 47.33 -> 62.67 | Improved |
| select with SHIFT instead of Q-PPL, 2% budget | RLVR data selection (Qwen2.5-Math-1.5B) · AMC | Pass@1 accuracy: 30.12 -> 38.55 | Improved |
| increase budget from 2% to 4% with SHIFT | SHIFT (Qwen2.5-Math-1.5B) · MATH-500 / AMC | Pass@1 accuracy: 62.67 -> 64.67 (MATH-500); 38.55 -> 36.14 (AMC) | Mixed |
| use SHIFT instead of Cluster at 4% budget | RLVR data selection (Qwen2.5-Math-1.5B) · AMC | Pass@1 accuracy: 40.96 -> 36.14 | Hurt |
| SHIFT with 2% data vs full-data RLVR | SHIFT (Qwen2.5-Math-1.5B) · AMC | Pass@1 accuracy: 33.73 -> 38.55 | Improved |
| select with SHIFT instead of Random, 0.1% budget | RLVR data selection (Qwen3-1.7B) · MedQA | Pass@1 accuracy: 45.28 -> 50.35 | Improved |
| select with SHIFT instead of Random, 0.1% budget | RLVR data selection (Qwen3-1.7B) · MedXpertQA-R/U | Pass@1 accuracy: 6.84 -> 9.62 (R); 8.56 -> 12.22 (U) | Improved |
| select with SHIFT instead of Random, 0.2% budget | RLVR data selection (Qwen3-1.7B) · PubMedQA | Pass@1 accuracy: 70.72 -> 70.80 | No clear effect |
| select with SHIFT instead of Random, Llama-3-8B, 2% budget | RLVR data selection (Llama-3-8B) · MATH-500 | Pass@1: 23.41 -> 25.17 | Improved |
| select with SHIFT instead of Random, OLMo-3-7B, 2% budget | RLVR data selection (OLMo-3-7B) · MATH-500 | Pass@1: 81.67 -> 83.67 | Improved |
| use RIRS-based coverage space phi_{s+Delta} for farthest-first instead of question embedding (add RIRS coverage) (ablation) | CoreSet selection (Qwen3-1.7B) · PubMedQA, 0.1% budget | accuracy: 68.60 -> 73.00 | Improved |
Closest papers by meaning
- Smart Picks in the Dark: Towards Efficient RLVR for Reasoning via Tracing Metacognitive Pivots (NeurIPS 2026 · similarity 0.91)
- Detecting RLVR Training Data via Structural Convergence of Reasoning (NeurIPS 2026 · similarity 0.90)
- On the Direction of RLVR Updates for LLM Reasoning: Identification and Exploitation (ICLR 2026 · similarity 0.89)
- What are Key Factors for Updates in RL for LLM Reasoning? (NeurIPS 2026 · similarity 0.89)
- Generalization of RLVR Using Causal Reasoning as a Testbed (ICLR 2026 · similarity 0.88)
- TraPO: A Semi-Supervised Reinforcement Learning Framework for Boosting LLM Reasoning (ICLR 2026 · similarity 0.88)
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.
