Paper 2603.04597
Bootstrapping Exploration with Group-Level Natural Language Feedback in Reinforcement Learning
- Published
- Mar 2026
- Research lab
- Independent
- Citations
- 2
- GitHub
- 18 stars
01 In brief
Summary
The paper introduces GOLF, a reinforcement learning (RL) framework that leverages group-level natural language (NL) feedback to improve exploration efficiency in training large language models (LLMs).
Unlike standard RL that relies solely on scalar rewards, GOLF aggregates two complementary feedback sources: external critiques (pinpointing errors or suggesting fixes) and intra-group attempts (alternative responses within a rollout group).
These are combined to generate high-quality refinements, which are adaptively injected as off-policy scaffolds when the policy struggles in sparse-reward regions.
GOLF jointly optimizes generation and refinement in a unified RL loop, creating a virtuous cycle that enhances both capabilities.
Experiments on non-verifiable benchmarks (e.g., AlpacaEval, WildBench) show GOLF outperforms the strongest baseline by 22.7% and achieves 2.2× sample efficiency over scalar-reward RL.
On verifiable tasks (math, instruction following, code generation), GOLF consistently improves performance and Pass@k, indicating broader solution coverage.
Ablations confirm the complementarity of feedback sources and the benefits of adaptive injection and joint optimization, with GOLF also improving self-refinement capabilities on RefineBench.
The code is available at https://github.com/LuckyyySTA/GOLF.
02 From the paper
Abstract
Large language models (LLMs) typically receive diverse natural language (NL) feedback through interaction with the environment. However, current reinforcement learning (RL) algorithms rely solely on scalar rewards, leaving the rich information in NL feedback underutilized and leading to inefficient exploration. In this work, we propose GOLF, an RL framework that explicitly exploits group-level language feedback to guide targeted exploration through actionable refinements. GOLF aggregates two complementary feedback sources: (i) external critiques that pinpoint errors or propose targeted fixes, and (ii) intra-group attempts that supply alternative partial ideas and diverse failure patterns. These group-level feedbacks are aggregated to produce high-quality refinements, which are adaptively injected into training as off-policy scaffolds to provide targeted guidance in sparse-reward regions. Meanwhile, GOLF jointly optimizes generation and refinement within a unified RL loop, creating a virtuous cycle that continuously improves both capabilities. Experiments on both verifiable and non-verifiable benchmarks show that GOLF achieves superior performance and exploration efficiency, achieving 2.2$\times$ improvements in sample efficiency compared to RL methods trained solely on scalar rewards. Code is available at https://github.com/LuckyyySTA/GOLF.