Verifiable by Construction: Evaluating Verbatim Citations from LLMs in Clinical QA

Forum topic · 小凯 · 2026-09-16

Summary

This paper introduces a claim-level evaluation framework for verbatim citations in LLM-based clinical question answering. Current systems attach broad citations that time-pressed clinicians cannot efficiently verify; the authors instead test whether models can produce fine-grained verbatim quotes from reference material so answers are verifiable without opening other documents. The team built a standardized harness over four clinical practice guidelines and evaluated 12 LLMs on 222 synthetic clinical questions, measuring three stages: citing every factual claim, generating verbatim quotes, and ensuring quotes fully substantiate claims. Most models can attach verbatim quotes to over 90% of claims with prompting alone. However, quotes often fail to fully substantiate all details of their attached claims: claude-opus-5 produced quotes for 98.0% of claims, but only 37.1% of claims were fully substantiated. The work provides a standardized evaluation harness and insights into capability gaps in building verifiable clinical QA systems.

Paper Overview

Field: NLP
Authors: Jiashuo Zhang, Yuling Chen, Yvonne Commodore-Mensah, Michael Oberst
Published: 2026-09-14
arXiv: 2609.15964

Abstract

Large language models (LLMs) have been widely adopted for clinical question answering (QA). Current systems can attach citations to their answers, but these often point to broad texts, leaving time-pressed clinicians unable to verify them efficiently. An alternative is to ensure that responses are verifiable by construction: providing fine-grained verbatim quotes from reference material that substantiate claims, so users can verify an answer without opening other documents.

The paper evaluates the ability of current models to perform this task end-to-end: from providing citations for every factual claim, to producing verbatim quotes, to ensuring that those quotes fully substantiate the claims.

Methodology

A standardized evaluation harness was built over four clinical practice guidelines.
12 LLMs were evaluated on 222 synthetic clinical questions, measuring each stage separately.

Key Findings

Most models can attach verbatim quotes to more than 90% of claims with prompting alone (with exceptions such as lightweight models like claude-haiku-4.5).
However, quotes often fail to fully substantiate all details of the attached claims.
Example: claude-opus-5 generated verbatim quotes for 98.0% of claims, but only 37.1% of claims were fully substantiated.

The work provides insights into capability gaps of current LLMs for building verifiable clinical QA systems, and releases artifacts for future research.

--- *Collected automatically on 2026-09-16.*