Aligning Clinical Needs and AI Capabilities: A Survey on LLMs for Medical Reasoning
06:00 · July 11, 2026 · arXiv cs.AI RSS
Large language models (LLMs) have emerged as important tools in healthcare, showing growing potential for clinical reasoning and patient care. This survey examines recent progress in medical LLMs, focusing on reasoning applications and requirements. We present a dual-view approach that connects clinical practice with computational methods. On the clinical side, we establish a five-level competency scheme following Miller's Pyramid, progressing from knowledge recall to dynamic case management. On the computational side, we link deductive, inductive, and abductive reasoning patterns to common medical goals and tasks. We also introduce a benchmark dataset spanning five levels of medical reasoning capability and report results on 18 state-of-the-art models, revealing that medical specialist models excel in diagnosis-centric tasks while general models lead in decision support and dialogue. We conclude by discussing current progress and open challenges, including data limitations, hallucination, and grounding issues, and outline directions toward safer, more reliable, and workflow-ready systems.
Summary
This survey examines how large language models can support clinical reasoning across the care pathway, from initial triage through diagnosis, treatment planning, and ongoing case management. Rather than treating medical LLMs as general text processors, the authors introduce a dual-view framework that aligns computational reasoning patterns with clinical requirements. On the clinical side, they extend Miller’s Pyramid into five progressive competency levels, ranging from basic knowledge recognition to dynamic, context-sensitive decision making under uncertainty. On the computational side, they map deductive, inductive, and abductive reasoning approaches to concrete medical tasks such as symptom normalization, risk prediction, differential diagnosis, and individualized treatment selection.
To enable consistent evaluation, the authors release a new benchmark containing 5,000 curated samples evenly distributed across the five competency levels. Systematic testing of 18 state-of-the-art models reveals clear specialization patterns: medical-domain models tend to outperform on diagnosis-focused tasks, whereas larger general-purpose models show advantages in multi-turn dialogue, summarization, and decision-support scenarios. Structured temporal predictions, such as length-of-stay estimation, remain difficult for all systems evaluated. The results indicate that factors beyond raw model size—instruction quality, domain-specific data, and explicit reasoning training—strongly influence outcomes.
The survey also reviews current training and inference techniques, including instruction tuning on expert data, chain-of-thought prompting, retrieval-augmented generation, tool-augmented agent workflows, and multimodal integration of text with imaging. Evaluation extends beyond accuracy to consider factuality, reasoning completeness, evidence grounding, and uncertainty calibration. Persistent challenges include limited high-quality training data, hallucination risks, and difficulties in reliably anchoring model outputs to verifiable clinical sources. The authors conclude by outlining practical deployment strategies, such as routing diagnosis-heavy queries to specialist models while using general models for dialogue and supportive tasks, and by identifying concrete directions for safer, workflow-integrated systems.
Why it matters
This survey provides a rigorous, structured framework for evaluating medical LLMs, which is highly valuable for Dutch AI researchers and healthcare institutions developing transparent and safe clinical AI. Its focus on mitigating hallucinations and ensuring reliable reasoning aligns well with the EU AI Act and the Netherlands' emphasis on ethical AI deployment.
