StepHint: Multi-level Stepwise Hints Enhance Reinforcement Learning to Reason

2025-07-03

AI Summary

To address the “near-miss penalty” (where minor errors invalidate otherwise correct reasoning chains) and “exploration stagnation” (where policies get trapped in local optima with insufficient global exploration) in large language models (LLMs) for complex reasoning, this paper proposes StepHint—a reinforcement learning framework leveraging verifiable rewards. StepHint introduces a novel multi-level stepwise prompting mechanism and an adaptive step partitioning strategy, hierarchically guiding the model to generate intermediate reasoning steps that are individually verifiable. This design alleviates reward sparsity and reduces reliance on locally optimal solutions. Evaluated on six mathematical reasoning benchmarks, StepHint significantly outperforms existing RLVR methods, demonstrating superior generalization. Moreover, it achieves substantial improvements over baseline models across cross-domain tasks, validating its enhanced exploration efficiency and reasoning robustness.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Reinforcement learning with verifiable rewards (RLVR) is a promising approach for improving the complex reasoning abilities of large language models (LLMs). However, current RLVR methods face two significant challenges: the near-miss reward problem, where a small mistake can invalidate an otherwise correct reasoning process, greatly hindering training efficiency; and exploration stagnation, where models tend to focus on solutions within their ``comfort zone,'' lacking the motivation to explore potentially more effective alternatives. To address these challenges, we propose StepHint, a novel RLVR algorithm that utilizes multi-level stepwise hints to help models explore the solution space more effectively. StepHint generates valid reasoning chains from stronger models and partitions these chains into reasoning steps using our proposed adaptive partitioning method. The initial few steps are used as hints, and simultaneously, multiple-level hints (each comprising a different number of steps) are provided to the model. This approach directs the model's exploration toward a promising solution subspace while preserving its flexibility for independent exploration. By providing hints, StepHint mitigates the near-miss reward problem, thereby improving training efficiency. Additionally, the external reasoning pathways help the model develop better reasoning abilities, enabling it to move beyond its ``comfort zone'' and mitigate exploration stagnation. StepHint outperforms competitive RLVR enhancement methods across six mathematical benchmarks, while also demonstrating superior generalization and excelling over baselines on out-of-domain benchmarks.

Problem

• Addresses near-miss rewards in RLVR for LLM reasoning
• Mitigates exploration stagnation in reinforcement learning models
• Enhances training efficiency with multi-level stepwise hints

Innovation

• Multi-level stepwise hints guide exploration
• Adaptive partitioning of reasoning chains
• Mitigates near-miss rewards and stagnation