Back to Basics: Revisiting Exploration in Reinforcement Learning for LLM Reasoning via Generative Probabilities
This addresses the problem of low-entropy policies in RL for LLM reasoning, which is incremental but important for improving diversity in mathematical and coding tasks.
The paper tackles the problem of mode collapse and limited output diversity in reinforcement learning for LLM reasoning by proposing an Advantage Re-weighting Mechanism (ARM) that equilibrates confidence levels across correct responses, resulting in improved generative diversity and competitive accuracy with gains of 5.7% in Pass@1 and 13.9% in Pass@32 on Qwen2.5-7B.
Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as an indispensable paradigm for enhancing reasoning in Large Language Models (LLMs). However, standard policy optimization methods, such as Group Relative Policy Optimization (GRPO), often converge to low-entropy policies, leading to severe mode collapse and limited output diversity. We analyze this issue from the perspective of sampling probability dynamics, identifying that the standard objective disproportionately reinforces the highest-likelihood paths, thereby suppressing valid alternative reasoning chains. To address this, we propose a novel Advantage Re-weighting Mechanism (ARM) designed to equilibrate the confidence levels across all correct responses. By incorporating Prompt Perplexity and Answer Confidence into the advantage estimation, our method dynamically reshapes the reward signal to attenuate the gradient updates of over-confident reasoning paths, while redistributing probability mass toward under-explored correct solutions. Empirical results demonstrate that our approach significantly enhances generative diversity and response entropy while maintaining competitive accuracy, effectively achieving a superior trade-off between exploration and exploitation in reasoning tasks. Empirical results on Qwen2.5 and DeepSeek models across mathematical and coding benchmarks show that ProGRPO significantly mitigates entropy collapse. Specifically, on Qwen2.5-7B, our method outperforms GRPO by 5.7% in Pass@1 and, notably, by 13.9% in Pass@32, highlighting its superior capability in generating diverse correct reasoning paths.