First Return, Entropy-Eliciting Explore

TL;DR

FR3E is a structured exploration framework for reinforcement learning in large language models that enhances reasoning stability, coherence, and correctness by targeting high-uncertainty decision points with semantically grounded feedback.

Contribution

FR3E introduces a novel exploration method that identifies high-uncertainty points and performs targeted rollouts to improve reasoning in LLMs without dense supervision.

Findings

Promotes more stable training in LLM reasoning.

Produces longer, more coherent responses.

Increases fully correct reasoning trajectories.

Abstract

Reinforcement Learning from Verifiable Rewards (RLVR) improves the reasoning abilities of Large Language Models (LLMs) but it struggles with unstable exploration. We propose FR3E (First Return, Entropy-Eliciting Explore), a structured exploration framework that identifies high-uncertainty decision points in reasoning trajectories and performs targeted rollouts to construct semantically grounded intermediate feedback. Our method provides targeted guidance without relying on dense supervision. Empirical results on mathematical reasoning benchmarks(AIME24) show that FR3E promotes more stable training, produces longer and more coherent responses, and increases the proportion of fully correct trajectories. These results highlight the framework's effectiveness in improving LLM reasoning through more robust and structured exploration.