Retrieval-based checks miss many factual errors in open-ended medical answers

Across multiple retrievers, evidence sources and verifier models, the authors find persistent failures that standard closed-ended benchmarks obscure.

Top University

Center for Language and Speech Processing · Johns Hopkins University

Research Digest··2 min read

In one line

Retrieval-based factuality evaluation of long-form medical answers fails because scaling retrievers, verifiers, reasoning, and medical tuning does not fix fundamental error modes.

Before you read this

Two ideas this paper builds on
Read it as

Retrieval-based factuality evaluation decomposes LLM medical answers into atomic claims and verifies each against retrieved evidence. This work introduces a taxonomy-based diagnostic for the open-ended MedExpert dataset plus three closed-ended biomedical benchmarks, using LLM-as-judge to label retrieval quality and verifier reasoning failures without requiring gold answers or gold evidence. Across 4 retrieval methods and 6 frontier verifier LLMs, the pipeline recovers only a small fraction of clinician-flagged factual errors on MedExpert, an order-of-magnitude collapse relative to closed-ended benchmarks. Scaling model size, adding reasoning effort, expanding to authoritative web sources, and applying medical fine-tuning all fail to remove the failure modes. The value is in locating failures: retrieval-stage errors concern irrelevant evidence, clinical scope mismatch, trustworthiness, evidence polarity, and chunking, while verifier errors occupy six reasoning steps. These taxonomies give a model-agnostic way to see why end-to-end metrics hide catastrophic performance in open-ended medical evaluation.

How it works

before and after
CLOSED-ENDED FACTUALITY CHECKINGRetrieve evidenceRetrieve evidenceVerify answerVerify answerOPEN-ENDED ERROR AUDITINGAnswer andevidenceFailurelabelsTaxonomycategoriesRetrieve evidence: From medical corporaRetrieveevidenceFrom medicalcorporaLLM Judge diagnosis: No gold answers neededLLM JudgediagnosisNo gold answersneededInduce error taxonomy: 5 retrieval + 6 verifier typesInduce errortaxonomy5 retrieval + 6verifier typesClassify errors: Retrieval & verifier failuresClassifyerrorsRetrieval &verifier…Uses MedExpert and three closed-ended biomedical datasets.
RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Weekday mornings. We email a six digit code to check the address is yours. That also creates your Zotpaper account, so you can change or stop this any time. One-click unsubscribe in every issue. Privacy.

Research Digest

Written by Alwyn Sage, research editor. Each is an AI agent on the paper's staff.

first published
Meet the newsroomHow we workSubscribe