flaviomerenda/fine-grained-medical-reasoning

Dataset Card for Fine-Grained Medical Reasoning

Fine-grained medical reasoning QA dataset introduced in "Can LLMs Reason Like Doctors? Exploring the Limits of Large Language Models in Complex Medical Reasoning" (Findings of EACL 2026). Manually annotated from the MedAgentsBench test_hard set, it evaluates LLMs’ abduction, deduction, and induction capabilities, offering detailed insights into physician-like reasoning.

Dataset Details

Dataset Description

• Curated by: Expert.ai Language Technology Research Lab
• Language: en
• License: MIT

Dataset Sources

 

Dataset Structure

• id: Source dataset ID
• dataset_name: Name of the source dataset
• reasoning_strategy: Reasoning approach required to answer the question
• question: Complex medical question from the MedAgentsBench test_hard set
• options: Possible answers to the question
• answer: Correct answer
• options_idx: Original indices of the answer options
• answer_idx: Index of the correct answer

Dataset Creation

Curation Rationale

LLMs may act primarily as advanced information retrievers rather than systems capable of genuine medical reasoning, as illustrated by the comparison below between a physician and llama-3.3-70b-instruct. Current LLMs may be unreliable when solving complex problems involving multi-step reasoning processes, such as abduction and deduction

Source Data

MedAgentsBench is a comprehensive benchmark that aggregates challenging subsets from eight widely used datasets, targeting advanced knowledge and multi-step reasoning.

Benchmark	Size	Avg Lens	Options	Description
MedBullets	89	210	5	USMLE Step 2 & 3 questions from an online platform requiring clinical reasoning
MedMCQA	100	17	4	Questions from real-world entrance exams AIIMS & NEET PG
MedQA	100	186	4	Medical board exam questions requiring clinical reasoning and decision-making
MedXpert-R	100	270	10	Advanced medical questions designed to assess expert-level reasoning capabilities
MedXpert-U	100	199	10	Advanced medical questions designed to assess expert-level knowledge
MMLU	73	51	4	Biomedical section of a multitask dataset spanning multiple academic domains
MMLU-Pro	100	46	10	Extends MMLU with harder reasoning-based questions and expands choices from 4 to 10
PubMedQA	100	326	3	Biomedical research questions from PubMed abstracts with yes/no/maybe answers

Annotations

Annotation process

Guided by the literature on medical problem-solving, a manually annotated subset has been created for fine-grained analysis of abductive, deductive, and inductive reasoning. Tasks selected for annotation correspond to those that existing literature associates with each reasoning type. Some tasks may involve chains of reasoning, such as abduction followed by deduction, which are annotated to capture the complete reasoning sequence.

The following table reports the annotation statistics for the resulting dataset.

Dataset	Abduction-Deduction	Abduction	Deduction	Induction
MedBullets	27	7	13	0
MedQA	3	4	2	0
MedXpertQA-R	18	9	5	0
MedXpertQA-U	0	2	0	0
MMLU	0	2	5	0
MMLU-PRO	2	1	0	0
PubMedQA	0	0	0	50
Total	50	25	25	50

Citation

BibTeX:

@inproceedings{merenda-etal-2026-llms,
    title = "Can {LLM}s Reason Like Doctors? Exploring the Limits of Large Language Models in Complex Medical Reasoning",
    author = "Merenda, Flavio  and
      Gomez-Perez, Jose Manuel  and
      Rigau, German",
    editor = "Demberg, Vera  and
      Inui, Kentaro  and
      Marquez, Llu{\'i}s",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EACL} 2026",
    month = mar,
    year = "2026",
    address = "Rabat, Morocco",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-eacl.127/",
    pages = "2432--2452",
    ISBN = "979-8-89176-386-9",
    abstract = "Large language models (LLMs) have shown remarkable progress in reasoning across multiple domains. However, it remains unclear whether their abilities reflect genuine reasoning or sophisticated pattern matching, a distinction critical in medical decision-making, where reliable multi-step problem-solving is required. Accordingly, we conduct one of the largest evaluations to date, assessing 77 LLMs with diverse fine-tuning approaches, ranging from 1 billion parameters to frontier models. Guided by medical problem-solving theory, we select three medical question answering (QA) benchmarks targeting key reasoning skills: reasoning processes, susceptibility to cognitive biases, and metacognitive abilities. Additionally, we manually annotate a subset of questions to assess the abduction, deduction, and induction capabilities of LLMs, offering detailed insight into the reasoning mechanisms followed by physicians, an aspect that has received relatively limited attention in this domain. Most models, particularly smaller ones, struggle even with specialized fine-tuning or advanced prompting. Larger models perform better but still show clear limitations in complex medical reasoning. Our findings highlight the need to improve specific reasoning strategies to better reflect medical decision-making. The datasets and code used in this study are publicly available at: \url{https://github.com/expertailab/Can-LLMs-Reason-Like-Doctors}"
}

Dataset Card Authors

Flavio Merenda

Dataset Card Contact

fmerenda@expert.ai