MedConclusion: A Benchmark for Biomedical Conclusion Generation from Structured Abstracts

Hugging Face Daily Papers Papers

Summary

MedConclusion introduces a large-scale benchmark of 5.7 million PubMed structured abstracts for evaluating LLMs on biomedical conclusion generation from structured scientific evidence. The study finds that conclusion writing is behaviorally distinct from summarization and that current automatic metrics cluster strong models closely together.

Large language models (LLMs) are widely explored for reasoning-intensive research tasks, yet resources for testing whether they can infer scientific conclusions from structured biomedical evidence remain limited. We introduce MedConclusion, a large-scale dataset of 5.7M PubMed structured abstracts for biomedical conclusion generation. Each instance pairs the non-conclusion sections of an abstract with the original author-written conclusion, providing naturally occurring supervision for evidence-to-conclusion reasoning. MedConclusion also includes journal-level metadata such as biomedical category and SJR, enabling subgroup analysis across biomedical domains. As an initial study, we evaluate diverse LLMs under conclusion and summary prompting settings and score outputs with both reference-based metrics and LLM-as-a-judge. We find that conclusion writing is behaviorally distinct from summary writing, strong models remain closely clustered under current automatic metrics, and judge identity can substantially shift absolute scores. MedConclusion provides a reusable data resource for studying scientific evidence-to-conclusion reasoning. Our code and data are available at: https://github.com/Harvard-AI-and-Robotics-Lab/MedConclusion.
Original Article
View Cached Full Text

Cached at: 04/21/26, 07:20 AM

Paper page - MedConclusion: A Benchmark for Biomedical Conclusion Generation from Structured Abstracts

Source: https://huggingface.co/papers/2604.06505

Abstract

A large-scale dataset of 5.7 million PubMed structured abstracts is introduced for biomedical conclusion generation, enabling evaluation of large language models’ ability to reason from structured scientific evidence.

Large language models(LLMs) are widely explored for reasoning-intensive research tasks, yet resources for testing whether they can infer scientific conclusions from structured biomedical evidence remain limited. We introduce MedConclusion, a large-scale dataset of 5.7M PubMedstructured abstractsforbiomedical conclusion generation. Each instance pairs the non-conclusion sections of an abstract with the original author-written conclusion, providing naturally occurring supervision forevidence-to-conclusion reasoning. MedConclusion also includes journal-level metadata such as biomedical category and SJR, enabling subgroup analysis across biomedical domains. As an initial study, we evaluate diverse LLMs under conclusion and summary prompting settings and score outputs with bothreference-based metricsandLLM-as-a-judge. We find that conclusion writing is behaviorally distinct from summary writing, strong models remain closely clustered under current automatic metrics, and judge identity can substantially shift absolute scores. MedConclusion provides a reusable data resource for studying scientificevidence-to-conclusion reasoning. Our code and data are available at: https://github.com/Harvard-AI-and-Robotics-Lab/MedConclusion.

View arXiv pageView PDFGitHub1Add to collection

Get this paper in your agent:

hf papers read 2604\.06505

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2604.06505 in a model README.md to link it from this page.

Datasets citing this paper2

#### harvardairobotics/MedConclusion-Compact Viewer• Updated1 day ago • 140k • 1 #### harvardairobotics/MedConclusion Viewer• Updated1 day ago • 5.83M • 2

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2604.06505 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Can AI Agents Synthesize Scientific Conclusions?

arXiv cs.AI

This paper introduces SciConBench, a large-scale benchmark with 9.11K questions and expert-written conclusions for evaluating AI agents' ability to synthesize scientific conclusions from open-domain evidence. The study finds that even the best agent achieves only a factual F1 of 0.337 in clean-room settings, highlighting that reliable synthesis remains an open challenge.

MedicalBench: Evaluating Large Language Models Toward Improved Medical Concept Extraction

arXiv cs.CL

MedicalBench is a new benchmark for evaluating large language models on medical concept extraction from electronic health records, focusing on implicit reasoning and evidence grounding. It includes 823 expert-annotated examples and shows that current models perform modestly, highlighting the difficulty of extracting implicitly stated medical concepts.

MedAction: Towards Active Multi-turn Clinical Diagnostic LLMs

arXiv cs.CL

This paper introduces MedAction, a framework for training LLMs on active, multi-turn clinical diagnosis by simulating iterative test ordering and hypothesis updates. It presents a new dataset, MedAction-32K, and demonstrates state-of-the-art performance for open-source models on medical benchmarks.

CalBrief: A Pilot Diagnostic Benchmark for Evidence-Calibrated Scientific Briefing with Large Language Models

arXiv cs.CL

The paper presents CalBrief, a pilot diagnostic benchmark of 16 evidence packages and 96 human-verified takeaways for evaluating whether large language models can generate evidence-calibrated scientific briefings. The study finds that structured organization improves reasoning but explicit strength-calibration policies are overly conservative, with most conservatism arising from expanded label spaces rather than signal injection.