Research map / Large language models / Reasoning and chain of thought

Hint Tuning: Less Data Makes Better Reasoners

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Hint Tuning uses an instruct model as a difficulty probe to label ~1K s1 problems into No-Hint, Sparse-Hint, and Full-Hint training traces, then SFTs reasoning models to calibrate reasoning length, cutting tokens 24-66% on math.

Problem

Large reasoning models generate verbose chain-of-thought uniformly regardless of problem difficulty, and existing efficiency methods either need large distillation sets or expensive RL with length penalties.

Approach

Use the corresponding instruct model as a capability probe: find the Minimum Effective Hint (smallest prefix of reasoning episodes that lets the instruct model answer correctly) to automatically label each problem into three reasoning-depth states, then fine-tune the reasoning model on the resulting variable-length traces with plain SFT.

Key results

• Hint Tuning reduces tokens 24-66% (31.5% average) across Qwen3-4B-Thinking and DeepSeek-R1-Distill 7B/14B/32B with 1K samples, e.g., Qwen3-4B 66% token reduction on MATH500 at -2.1 accuracy.
• On DeepSeek-R1-Distill-Qwen-32B, MATH500 accuracy rises 1.1 points (94.8 to 95.9) while tokens drop 47%.
• On cross-domain tasks, token use does not drop and sometimes rises (e.g., +41% on LiveCodeBench for DS-R1-7B) while accuracy is roughly preserved.

Surprising findings

• Self-generated traces outperform traces from the stronger DeepSeek-R1 teacher, and dropping incorrect-answer traces hurts accuracy.
• Compression does not transfer to OOD domains: token counts increase on GPQA-Diamond and LiveCodeBench for some models.

Limitations

• Relies on a capability gap between instruct and reasoning models; calibration across families and scales is open.
• Experiments focus on mathematical reasoning; generalization to other domains is limited and compression can disappear OOD.
• Mechanism by which structural episode tags induce adaptive reasoning is empirical only.

Experimental setup

Datasets: s1K AIME24 AIME25 HMMT25 MATH500 AMO-Bench LiveCodeBench GPQA-Diamond

Benchmarks: MATH500 AIME24 AIME25 HMMT25 AMO-Bench LiveCodeBench GPQA-Diamond

Metrics: accuracy token count

Baselines: Budget Prompt SFT-S1.1 VeriThinker EfficientR AdaptThink

Builds on: s1 VeriThinker AdaptThink EfficientR

Closest papers by meaning

• DiffAdapt: Difficulty-Adaptive Reasoning for Token-Efficient LLM Inference (ICLR 2026 · similarity 0.90)
• LYNX: Learning Dynamic Exits for Confidence-Controlled Reasoning (NeurIPS 2026 · similarity 0.90)
• Efficient Thought Space Exploration Through Strategic Intervention (AAAI 2026 · similarity 0.89)
• Stop When Further Reasoning Won’t Help: Attention-State Adaptive Generation in Reasoning Models (ICML 2026 · similarity 0.89)
• How Far Are We from Optimal Reasoning Efficiency? (NeurIPS 2025 · similarity 0.89)
• Reasoning-Trace Collapse: Evaluating the Loss of Explicit Reasoning During Fine-Tuning (NeurIPS 2026 · similarity 0.89)