Where Does Retrieval-Based Open-Ended Evaluation Fail? Automatic Taxonomy Induction from Long-Form Medical Answer Factuality Verification
A medical RAG study finds retrieve-then-verify factuality checks fail for structural reasons, not weak models.
The paper studies where retrieval-based factuality evaluation fails for open-ended medical answers, using MedExpert plus three closed-ended datasets. It defines taxonomies for retrieval-stage errors across five quality dimensions and verifier-reasoning errors across six steps, labeled at scale with an LLM-as-Judge pipeline. Stress tests over four retrieval methods and six frontier verifier models find that larger models, extra reasoning, authoritative web sources, and medical fine-tuning do not remove the failures. The authors argue these are structural limits of retrieve-then-verify evaluation and release code and data.
- Study targets open-ended medical factuality checks where gold answers and evidence are unavailable.
- Two taxonomies cover five retrieval-quality dimensions and six verifier-reasoning steps.
- LLM-as-Judge labels evidence quality and reasoning errors at scale.
- Tested across four retrieval methods and six frontier verifier models.
- Larger models, more reasoning, web sources, and medical fine-tuning do not fix the failures.
From the original article
1 min read at huggingface.co
Retrieval-based factuality evaluation, where LLM-generated claims are verified against evidence from authoritative medical corpora, has become the dominant paradigm for scalable hallucination detection in high-stakes clinical settings. Despite the urgency of reliable and transparent medical fact verification, most systems measure performance with aggregate metrics like F1, which obscure where and why failures occur.
By Heyuan Huang, Jirui Dai, Alexandra DeLucia · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.