Smaller Models are Natural Explorers for Policy-Level Diversity in GRPO

• URL: http://arxiv.org/abs/2605.30789v2
• Date: Tue, 02 Jun 2026 09:15:19 GMT
• Title: Smaller Models are Natural Explorers for Policy-Level Diversity in GRPO
• Abstract summary: Small models inherently exhibit higher policy-level diversity, indicated by their superior pass@k relative to larger counterparts as sample counts increase.<n>We propose S2L-PO (Small-to-Large Policy Optimization), a framework that leverages fixed small models as natural explorers to train larger models.<n>S2L-PO improves accuracy on diverse mathematical reasoning benchmarks (e.g., +8.8% on AIME 24 using a 1.7B explorer to guide the 8B model) while reducing rollout compute.
• Score: 80.3917290909154
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: We identify a new dimension for enhancing rollout diversity in Group Relative Policy Optimization (GRPO) for LLMs. While GRPO relies on diverse rollouts, prevailing strategies primarily increase diversity by injecting more token-level randomness, which may introduce step-wise noise and lead to incoherent trajectories. We uncover that smaller models within the same model family inherently exhibit higher policy-level diversity, indicated by their superior pass@k relative to larger counterparts as sample counts increase. Unlike token-level noise, this diversity is temporally correlated, preserves logical consistency, and provides structured exploration signals for gradient estimation. We thus propose S2L-PO (Small-to-Large Policy Optimization), a framework that leverages fixed small models as natural explorers to train larger models. To balance exploration and exploitation, we design a progressive annealing strategy that transitions from offline small-model rollouts to the large learner's own sampling. This shift elegantly avoids mid-training performance drops caused by the small model's capacity limits, achieving faster convergence and unlocking a higher performance ceiling. S2L-PO improves accuracy on diverse mathematical reasoning benchmarks (e.g., +8.8% on AIME 24 using a 1.7B explorer to guide the 8B model) while reducing rollout compute.

Related papers

• AdaGRPO: A Capability-Aware Adaptive Enhancement for Flow-based GRPO [78.36537400975298]
  Group Relative Policy Optimization (GRPO) has demonstrated remarkable success in aligning text-to-image (T2I) flow models with human preferences.<n>We have identified that the learning loop of current flow-based GRPO is fundamentally decoupled from the learner's current capability.<n>We propose Adaptive GRPO (AdaGRPO), a novel capability-aware RL algorithm tailored for flow models.
  arXiv  Detail & Related papers  (2026-06-05T02:07:08Z)
• Too Correct to Learn: Reinforcement Learning on Saturated Reasoning Data [55.84428098924793]
  We propose Constrained Uniform Top-K Sampling (CUTS), a parameter-free decoding strategy enforcing structure-preserving exploration.<n>We integrate this into Mixed-CUTS, a training framework synergizing exploitative and exploratory rollouts to amplify intra-group advantage variance.<n> Notably, Mixed-CUTS improves Pass@1 accuracy on the challenging AIME25 benchmark by up to 15.1% over standard GRPO.
  arXiv  Detail & Related papers  (2026-04-20T16:43:28Z)
• OpenVLThinkerV2: A Generalist Multimodal Reasoning Model for Multi-domain Visual Tasks [81.07598709704628]
  We present a novel RL training objective that replaces standard linear scaling with non-linear distributional matching.<n>We also present OpenVLThinkerV2, a highly robust, general-purpose multimodal model.
  arXiv  Detail & Related papers  (2026-04-09T17:59:39Z)
• Policy-based Tuning of Autoregressive Image Models with Instance- and Distribution-Level Rewards [16.135177543347773]
  Autoregressive (AR) models are highly effective for image generation, yet their standard maximum-likelihood estimation training lacks direct optimization for sample quality and diversity.<n>We propose a lightweight RL framework that casts token-based AR inference as a Markov Decision Process, optimized via Group Relative Policy Optimization.<n>Our core contribution is the introduction of a novel distribution-level Leave-One-Out FID (LOO-FID) reward.
  arXiv  Detail & Related papers  (2026-03-24T11:28:36Z)
• Adaptive Divergence Regularized Policy Optimization for Fine-tuning Generative Models [31.470613363668672]
  Adaptive Divergence Regularized Policy Optimization automatically adjusts regularization strength based on advantage estimates.<n>Our implementation with Wasserstein-2 regularization for flow matching generative models achieves remarkable results on text-to-image generation.<n> ADRPO generalizes to KL-regularized fine-tuning of both text-only LLMs and multi-modal reasoning models.
  arXiv  Detail & Related papers  (2025-10-20T19:46:02Z)
• G$^2$RPO: Granular GRPO for Precise Reward in Flow Models [74.21206048155669]
  We propose a novel Granular-GRPO (G$2$RPO) framework that achieves precise and comprehensive reward assessments of sampling directions.<n>We introduce a Multi-Granularity Advantage Integration module that aggregates advantages computed at multiple diffusion scales.<n>Our G$2$RPO significantly outperforms existing flow-based GRPO baselines.
  arXiv  Detail & Related papers  (2025-10-02T12:57:12Z)
• MMR1: Enhancing Multimodal Reasoning with Variance-Aware Sampling and Open Resources [113.33902847941941]
  Variance-Aware Sampling (VAS) is a data selection strategy guided by Variance Promotion Score (VPS)<n>We release large-scale, carefully curated resources containing 1.6M long CoT cold-start data and 15k RL QA pairs.<n> Experiments across mathematical reasoning benchmarks demonstrate the effectiveness of both the curated data and the proposed VAS.
  arXiv  Detail & Related papers  (2025-09-25T14:58:29Z)
• Stepwise Guided Policy Optimization: Coloring your Incorrect Reasoning in GRPO [22.00487909203855]
  Group Relative Policy Optimization fails to update a policy when all responses within a group are incorrect.<n>This limitation underscores a key gap between artificial and human intelligence.<n>We introduce a simple framework that mitigates the all-negative-sample issue by incorporating response diversity within groups.
  arXiv  Detail & Related papers  (2025-05-16T18:02:05Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.