A medical claim is only as correct
as the structure around it

Atom-level decomposition fragments clinical judgments, leaving the verifier with claims that are formally standalone but clinically incomplete. MedSNIP verifies snippets instead, clause-grouped units that keep the structure intact.

Hasan Iqbal¹, Sarfraz Ahmad¹, Hyunjae Kim², Sihyeon Park³, Junjie Liao^4,5, Qingyu Chen², Preslav Nakov¹, Yuxia Wang⁵

¹MBZUAI · ²Yale · ³Korea University · ⁴Beijing Normal · ⁵INSAIT, Sofia University

Paper Dataset Code

The problem, in one judgment

Watch a judgment come apart

Three real judgments from MedSNIP-Bench, each grouped by a different structural pattern. Toggle the verification unit and see what the verifier actually receives.

276

RESPONSES

2,524

SNIPPETS

PATTERNS

CORPORA

Three judgments, three fragmentations

The same failure, three ways

Three clinical judgments, each preserved as one snippet but split into three or four atoms — Each column is one judgment from a clinical-vignette answer. Snippet-level decomposition keeps the evidence, its interpretation, and the caveat as a single verifiable unit. Atom-level decomposition splits the same judgment into three or four claims.

Taxonomy

Six ways a judgment holds together

Annotators labelled every snippet with the structure that justifies it. Three patterns merge atoms and three keep a claim standalone.

Enumeration

1,008 snippets

Causal or conditional chain

225 snippets

Conclusion with its premises

321 snippets

Standalone claim

617 snippets

Distinct alternative

39 snippets

Deliberately isolated

314 snippets

A–C merge related atoms · D–F keep a claim atomic. The largest merge-pattern gain is on B, the causal-conditional chains.

Citation

@inproceedings{
anonymous2026medsnip,
title={Med{SNIP}: Building and Benchmarking Snippet-Level Granularity for Medical Fact Verification},
author={Anonymous},
booktitle={The 2026 Conference on Empirical Methods in Natural Language Processing},
year={2026},
url={https://openreview.net/forum?id=1vYUGLZnvr}
}