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

Hugging Face Daily Papers Papers

Summary

The S2L-PO framework uses smaller models as natural explorers to enhance policy diversity in GRPO for training large language models. It achieves faster convergence and improves accuracy on mathematical reasoning benchmarks while reducing rollout compute.

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.
Original Article
View Cached Full Text

Cached at: 06/15/26, 09:03 AM

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

Source: https://huggingface.co/papers/2605.30789 Authors:

,

,

,

,

,

,

,

,

,

Abstract

Small-to-Large Policy Optimization framework uses smaller models as natural explorers to enhance policy diversity and improve large language model training efficiency.

We identify a new dimension for enhancingrollout diversityinGroup Relative Policy Optimization(GRPO) for LLMs. While GRPO relies on diverse rollouts, prevailing strategies primarily increase diversity by injecting moretoken-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 higherpolicy-level diversity, indicated by their superiorpass@krelative to larger counterparts as sample counts increase. Unlike token-level noise, this diversity is temporally correlated, preserves logical consistency, and provides structured exploration signals forgradient 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 aprogressive annealingstrategy 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 diversemathematical reasoning benchmarks(e.g., +8.8% on AIME 24 using a 1.7B explorer to guide the 8B model) while reducing rollout compute.

View arXiv pageView PDFProject pageGitHub2Add to collection

Get this paper in your agent:

hf papers read 2605\.30789

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.30789 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.30789 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.30789 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

SLPO: Scaling Latent Reasoning via a Surrogate Policy

Hugging Face Daily Papers

Introduces Surrogate Latent Policy Optimization (SLPO) to apply outcome-reward RL to autoregressive latent reasoners, enabling test-time scaling and variable-horizon policies that improve accuracy on harder instances.