Research map / Reinforcement learning / RL from feedback
Smaller Models are Natural Explorers for Policy-Level Diversity in GRPO
Yiming Ren, Yiran Xu, Zicheng Lin, Chufan Shi, Yukang Chen, Dingdong WANG, Tianhe Wu, Junjie Wang, Yujiu Yang, Yu Qiao, Ruihang Chu
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 |
|---|---|---|---|
| sample part of each GRPO group from a frozen 1.7B model, annealed to on-policy | Qwen3-8B-Base GRPO · AIME24 | Pass@1 (%): 15.0 -> 23.8 | Improved |
| sample part of each GRPO group from a frozen 1.7B model, annealed to on-policy | Qwen3-8B-Base GRPO · AIME25 | Pass@1 (%): 12.1 -> 22.5 | Improved |
| sample part of each GRPO group from a frozen 4B model, annealed to on-policy | Qwen3-14B-Base GRPO · AIME24, AIME25, MATH-500, OlympiadBench | Pass@1 (%): AIME24 18.0 -> 24.4; AIME25 12.9 -> 14.6 | Improved |
| sample part of each GRPO group from a frozen 1.8B model, annealed to on-policy | InternLM2.5-7B GRPO · AIME24, AIME25, MATH-500, OlympiadBench | Pass@1 (%): MATH-500 18.6 -> 22.6 | Improved |
| S2L-PO training on math, evaluated out of domain | Qwen3-8B-Base · CommonsenseQA | accuracy (%): 63.9 -> 67.8 | Improved |
| S2L-PO with 1.7B explorer on out-of-domain | Qwen3-8B-Base · CommonsenseQA | accuracy (%): 63.9 -> 64.2 | No clear effect |
| S2L-PO-4B on 14B out of domain | Qwen3-14B-Base · CommonsenseQA | accuracy (%): 67.2 -> 70.7 | Improved |
| filter small-model rollouts so diversity matches the large model (ablation) | S2L-PO 1.7B->8B · AIME24 | Pass@1 (%): 23.8 -> 14.7 | Hurt |
| raise sampling temperature to 1.5 (token-level noise) instead of small-model rollouts (ablation) | GRPO · AIME24&25 | Pass@1 | Hurt |
| use small-only rollouts without transitioning to on-policy GRPO (ablation) | S2L-PO · AIME24&25 | Pass@1 | Hurt |
| use progressive linear annealing instead of abrupt two-phase switch (ablation) | S2L-PO · Qwen3-8B-Base AIME | Pass@1 | Improved |
| lengthen annealing from 5 to 8 steps (ablation) | S2L-PO · Qwen3-8B-Base AIME | Pass@1 | Improved |
Closest papers by meaning
- Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO (NeurIPS 2026 · similarity 0.90)
- SetPO: Set-Level Policy Optimization for Diversity-Preserving LLM Reasoning (ICML 2026 · similarity 0.89)
- WS-GRPO: Weakly-Supervised Group-Relative Policy Optimization for Rollout-Efficient Reasoning (ICML 2026 · similarity 0.89)
- Act Only When It Pays: Efficient Reinforcement Learning for LLM Reasoning via Selective Rollouts (NeurIPS 2025 · similarity 0.88)
- Diversity-Aware Policy Optimization for Large Language Model Reasoning (NeurIPS 2025 · similarity 0.88)
- Sample More to Think Less: Group Filtered Policy Optimization for Concise 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.
