Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Kaiyi Zhang, Ang Lv, Jinpeng Li, Yongbo Wang, Feng Wang, Haoyuan Hu, Rui Yan

TL;DR

StepHint introduces multi-level stepwise hints for RLVR to address near-miss rewards and exploration stagnation in LLM reasoning. It uses adaptive end-of-thinking-based partitioning to create meaningful reasoning steps from stronger-model chains, then provides multiple hint levels to guide learning while preserving exploration. The approach couples with GRPO (and with PPO variants) and includes a GRPO-specific adjustment to handle hint prefixes, achieving superior results on six math benchmarks and robust out-of-domain generalization. Empirical results show higher accuracy, improved pass@k performance, and richer training dynamics, indicating enhanced reasoning capabilities and generalization. The work suggests a scalable path to leverage external reasoning chains without sacrificing independent exploration in RLVR.

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.

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

TL;DR

StepHint introduces multi-level stepwise hints for RLVR to address near-miss rewards and exploration stagnation in LLM reasoning. It uses adaptive end-of-thinking-based partitioning to create meaningful reasoning steps from stronger-model chains, then provides multiple hint levels to guide learning while preserving exploration. The approach couples with GRPO (and with PPO variants) and includes a GRPO-specific adjustment to handle hint prefixes, achieving superior results on six math benchmarks and robust out-of-domain generalization. Empirical results show higher accuracy, improved pass@k performance, and richer training dynamics, indicating enhanced reasoning capabilities and generalization. The work suggests a scalable path to leverage external reasoning chains without sacrificing independent exploration in RLVR.

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.

Paper Structure

This paper contains 32 sections, 1 theorem, 11 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Step 1: Adaptive Stepwise Partitioning of Reasoning Chains.
  3. Step 2: Multi-Level Hints for Problem Solving.
  4. Background: Reinforcement Learning with Verifiable Rewards
  5. Policy optimization algorithms
  6. Proximal Policy Optimization (PPO)
  7. Group Relative Policy Optimization (GRPO)
  8. Method
  9. Preliminaries: A Solution Space View of Reasoning
  10. StepHint: Multi-level Stepwise Hints Enhance RLVR
  11. Adaptive Stepwise Partitioning of On-Hand Reasoning Chains
  12. Definitions
  13. Method Details
  14. Multi-level Hints for Problem Solving
  15. Experiments
  16. ...and 17 more sections

Key Result

Proposition 1

Let $\mathcal{R}$ be the solution space and $S_{k-1}$ be the state after $k-1$ tokens have been generated. Upon generating the next token $t_k$ to form state $S_k = (S_{k-1}, t_k)$, the expected entropy of the solution space is bounded by the current entropy:

Figures (4)

  • Figure 1: Adaptive stepwise partitioning of a reasoning chain: step boundaries are identified where the probability of concluding the reasoning chain after the current token, $p(\texttt{</think>}|G_i)$, is greater than concluding after the subsequent token, $p(\texttt{</think>}|G_{i+1})$. A final partition is chosen to meet constraints on step count $m$ and minimum length $l$.
  • Figure 2: An overview of the multi-level hinting process. The process begins with a ground-truth reasoning chain, which is partitioned into $m$ steps (Section \ref{['sec:seg']}). From these steps, we construct $m-1$ prefix-based hints$\left(h_1,h_2,\cdots,h_{m-1}\right)$. The model is trained to generate completions from each hint level, as well as from scratch (Unhinted), and a reference trajectory.
  • Figure 3: Comparison of pass@k results on the AIME24 and AIME25 datasets. Left: AIME24; Right: AIME25.
  • Figure 4: Training dynamics of StepHint compared with GRPO. Left: Reward; Middle: Entropy; Right: Response Length.

Theorems & Definitions (2)

  • Proposition 1
  • proof