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

Optimizing Inference-Time Compute for Medical Reasoning via Uncertainty Quantification

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking

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.

Proposes AdaThink-Med, an entropy-and-correctness difficulty-aware length reward within GRPO that lets a single medical LLM adaptively switch between short and long reasoning, cutting tokens 4.7–6.4× on six medical QA benchmarks.

Problem

Medical reasoning LLMs apply long chain-of-thought uniformly to simple and complex queries, wasting inference compute and latency; existing length-penalty RL methods ignore problem difficulty and can collapse.

Approach

Estimate per-question difficulty by combining rollout correctness with top-K token entropy as uncertainty, then use a piecewise length reward (penalize length on easy-correct, reward length on hard-incorrect, zero otherwise) with a performance-compensation term inside GRPO, yielding thinking and non-thinking modes in one model without a router.

Key results

• On Qwen2.5-7B, AdaThink-Med reduces average output length from 497 to 106 tokens (4.7x) with +0.25% accuracy change, AES 0.93.
• On Llama-3.1-8B, average length drops from 410 to 64 tokens (6.4x) with -1.13% accuracy, AES 0.92 vs 0.78 for Kimi1.5 (best baseline).
• Human evaluation by three physicians on 500 cases reports Logical Soundness of 86.40% vs 62.8-69.5% for length-reduction baselines.

Surprising findings

• Greedy length calibration (Kimi1.5, ShortBetter) collapses in accuracy and length during RL, attributed to length reward hacking.
• Llama backbones reward-hack via repetition on incorrect hard questions and require an n-gram repetition penalty; Qwen does not show this.

Limitations

• Relies on entropy quality; miscalibrated models or homogeneous rollouts may make entropy a weak difficulty proxy.
• Evaluation is primarily single-turn QA; multi-turn and long-context clinical workflows are untested.
• Llama backbones need an n-gram repetition penalty to avoid degenerate repetition; tau and alpha may need retuning across domains.

Experimental setup

Datasets: AlphaMed19k MedQA MedMCQA

Benchmarks: MedQA MedMCQA PubMedQA MMLU-ProM GPQA-M MedXpert

Metrics: accuracy output length (tokens) AES (accuracy-efficiency score)

Baselines: Qwen2.5-7B-Instruct Llama-3.1-8B-Instruct MedLlama3 MMed Med42 OpenBioLLM UltraMedical HuatuoGPT-o1 m1 Kimi1.5 ShortBetter DAST

Builds on: GRPO Kimi1.5 ShortBetter DAST HuatuoGPT-o1

Closest papers by meaning

• MedS³: Towards Medical Slow Thinking with Self-Evolved Soft Dual-sided Process Supervision (AAAI 2026 · similarity 0.88)
• Simulating Viva Voce Examinations to Evaluate Clinical Reasoning in Large Language Models (NeurIPS 2025 · similarity 0.86)
• Chiron-o1: Igniting Multimodal Large Language Models towards Generalizable Medical Reasoning via Mentor-Intern Collaborative Search (NeurIPS 2025 · similarity 0.85)
• OctoMed: Data Recipes for State-of-the-Art Multimodal Medical Reasoning (CVPR 2026 · similarity 0.85)
• MTBBench: A Multimodal Sequential Clinical Decision-Making Benchmark in Oncology (NeurIPS 2025 · similarity 0.85)
• KGARevion: An AI Agent for Knowledge-Intensive Biomedical QA (ICLR 2025 · similarity 0.85)