Hint Tuning: Less Data Makes Better Reasoners

2026-05-09

AI Summary

This work addresses the inefficiency of existing large reasoning models, which often generate excessively long chains of thought without adapting to problem difficulty, resulting in substantial token redundancy. To mitigate this, the authors propose Hint Tuning—a method that leverages an instruction-tuned model as a difficulty detector to automatically construct three types of training samples: no-hint, sparse-hint, and full-hint, based on the detector’s performance under varying prompts. Problem difficulty is encoded as a consistency check between the instruction model and the reasoning model. Using only 1K self-annotated samples, combined with trinary-prompt fine-tuning and lightweight alignment training, the approach reduces generated tokens by 31.5% on average (up to 66%) across mainstream reasoning models ranging from 4B to 32B parameters, while maintaining competitive accuracy on five standard benchmarks.

Abstract

Large reasoning models achieve high accuracy through extended chain-of-thought but generate 5--8 more tokens than necessary, applying verbose reasoning uniformly regardless of problem difficulty. We propose Hint Tuning, a data-efficient approach that teaches models to calibrate reasoning depth. Our key insight: the corresponding instruct model serves as an ideal difficulty probe. By testing what the instruct model can solve with varying guidance, we automatically construct training data across three states: No-Hint (direct answer), Sparse-Hint (minimal prefix), and Full-Hint (complete reasoning). This converts the abstract challenge of difficulty labeling into a measurable consistency check between the instruct and reasoning models. With only 1K self-annotated samples, Hint Tuning achieves 24--66% token reduction (31.5% average) across mainstream reasoning models (Qwen3-Thinking, DeepSeek-R1-Distill) at multiple scales (4B--32B) while maintaining competitive accuracy on five benchmarks. Unlike methods requiring massive distillation datasets or expensive RL, we achieve superior efficiency through simple alignment with the instruct model's capabilities.

Problem

• reasoning efficiency
• token reduction
• reasoning depth calibration
• chain-of-thought
• model verbosity

Innovation

• Hint Tuning
• reasoning efficiency
• difficulty-aware reasoning
• data-efficient alignment
• token reduction