When Retrieval Helps and Distracts: Evaluating Evidence-Generating LLMs for Biomedical Claim Verification

Retrieval isn't universally helpful for biomedical fact-checking: While retrieving scientific papers from PubMed improved performance on specialized sources like PubMedQA and SciFact, it actually distracted models when fact-checking broader public health claims, suggesting that one-size-fits-all retrieval strategies may be counterproductive.

Fine-tuned language models outperform retrieval-augmented systems for evidence generation: When the task requires not just predicting whether a claim is true or false, but also providing useful evidence to support that verdict, fine-tuned LLMs performed better than systems that retrieved external documents, indicating that specialized training may be more valuable than access to reference materials.

Current metrics for retrieval quality are insufficient: The paper found that traditional measures like retrieval recall and lexical overlap between retrieved and reference evidence don't adequately capture whether retrieved information actually helps models make correct decisions—highlighting a gap in how we evaluate biomedical AI systems.

Evidence quality matters as much as accuracy: The research emphasizes that biomedical fact-checking systems must produce evidence that is "faithful, complete, and useful," not just accurate verdicts, raising the bar for what responsible AI in healthcare should accomplish.

Source context determines optimal system design: The finding that retrieval utility is "source-dependent" suggests that effective biomedical fact-checking requires tailoring approaches to specific types of claims rather than deploying uniform solutions across all health-related verification tasks.