Technical Summary: When Retrieval Helps and Distracts

Problem Statement

Biomedical fact-checking systems face a dual challenge: they must not only predict whether a claim is supported, contradicted, or unaddressed but also generate evidence that is faithful, complete, and useful for human review. Current Retrieval-Augmented Generation (RAG) approaches often assume that retrieving authoritative biomedical abstracts (e.g., from PubMed) automatically improves verification. However, in heterogeneous benchmarks combining scientific claims with public health reporting and misinformation, retrieved documents may be topically related yet irrelevant to the specific verification decision. This can lead to "distracted" models that generate fluent but unfaithful evidence, potentially converting associations into causal claims or omitting critical uncertainty.

The paper investigates these issues using CARE-XAI, a unified benchmark spanning five sources (PubMedQA, SciFact, HealthVer, PUBHEALTH, and HealthFC) with 17,803 examples. The study aims to determine:

1. How verdict-only systems differ from evidence-generating systems.
2. When PubMed retrieval aids or hinders verification.
3. Whether the decision utility of retrieved evidence can be measured relative to trusted reference evidence.

Methodology

The authors conducted a systematic evaluation under a shared protocol comparing five system families:

• Base Instruction LLMs: Zero-shot and few-shot prompting (including Chain-of-Thought and PICO frameworks) on models like Gemma, Qwen, and Gemini.
• PubMed RAG LLMs: Base LLMs augmented with a frozen retrieval pipeline using Entrez for PubMed queries and BGE-M3 for reranking.
• Fine-tuned LLMs: Models (Gemma and Qwen) adapted via LoRA on the CARE-XAI training split to learn the specific evidence formatting and label semantics.
• Label-only LLMs & Biomedical Encoders: Systems optimized solely for verdict prediction (e.g., BioBERT, BioLinkBERT) to serve as diagnostic probes.

Evaluation Metrics:

• Verdict Quality: Macro-F1 (primary), accuracy, balanced accuracy, and MCC.
• Evidence Quality: Evidence coverage, lexical overlap (ROUGE/BERTScore as secondary diagnostics), and explanation-verdict consistency.
• Retrieval Diagnostics: Query success, MRR, Recall@k, and nDCG@10.
• Bio-GRACE: A novel, gold-reference-normalized diagnostic metric. It measures the "utility recovery" of retrieved context by comparing the probability shift a verifier makes when using retrieved evidence versus gold (reference) evidence. It quantifies how much of the decision benefit provided by trusted evidence is recovered by the retrieved context.

Experimental Rigor:

• A frozen retrieval cache was built for the test set to ensure retrieval quality was decoupled from generation performance.
• Leakage audits were performed to filter out rows with evidence overlap between train and test splits.
• Human Verification: 100 blinded outputs were evaluated by two biotechnology annotators and two LLM judges on verdict correctness, evidence support, usefulness, and safety.

Key Contributions

1. Unified Benchmark (CARE-XAI): A consolidated dataset integrating five distinct biomedical fact-checking sources, highlighting the heterogeneity of evidence needs (from scientific abstracts to policy/news).
2. Bio-GRACE Metric: A diagnostic tool that moves beyond simple retrieval recall to measure whether retrieved context actually recovers the decision-making benefit of reference evidence. It reveals that retrieval utility is highly source-dependent.
3. Systematic Comparison: A comprehensive evaluation separating verdict prediction from evidence generation, demonstrating that high-performing classifiers do not necessarily generate high-quality evidence.
4. Retrieval Utility Analysis: Evidence that "always-on" RAG is suboptimal for heterogeneous biomedical fact-checking, as it can degrade performance on public health claims while helping scientific abstract claims.

Key Results

• Verdict vs. Evidence Generation: Biomedical encoder classifiers remain the strongest performers for verdict-only prediction (Macro-F1 ~0.56), but they cannot generate evidence. Among evidence-generating systems, fine-tuned LLMs outperform base and RAG LLMs (Best Fine-tuned F1: 0.483 vs. Best Base/RAG F1: ~0.45).
• The Mixed Impact of Retrieval:
  • Helpful: PubMed retrieval significantly improves performance on PubMedQA and SciFact (sources aligned with scientific abstracts).
  • Distracting: Retrieval often degrades performance on PUBHEALTH and HealthFC (public health/misinformation claims). Bio-GRACE shows negative utility recovery for these sources, indicating that retrieved abstracts often distract the model from the correct decision.
• Fine-Tuning Superiority: Fine-tuning is a more reliable adaptation strategy for evidence-generating LLMs than RAG. It improves consistency across prompts and models without the volatility introduced by retrieval noise.
• Evidence Conditioning: When provided with gold (reference) evidence, both classifiers and LLMs show substantial improvements in verdict prediction, confirming the task is evidence-sensitive. However, retrieved PubMed evidence fails to close the gap to gold evidence, identifying retrieval/source matching as the primary bottleneck.
• Human Evaluation: Human agreement on evidence support and usefulness was modest (Cohen's κ ~0.39 and ~0.15, respectively), highlighting the difficulty of the task. Fine-tuned models showed the highest agreed verdict correctness, while RAG models showed higher agreed evidence support on specific subsets, though overall performance remained low.

Significance and Claims

The paper claims that biomedical verification requires treating evidence as a "first-class, decision-useful, auditable, and safety-critical output." The primary significance lies in demonstrating that:

1. Retrieval is not a panacea: Simply adding PubMed retrieval to LLMs does not guarantee better fact-checking; in heterogeneous settings, it can introduce noise that distracts the model.
2. Source-Dependent Utility: The utility of retrieval is strictly tied to the source of the claim. A "one-size-fits-all" retrieval strategy is insufficient; selective retrieval (gating retrieval based on source type or model uncertainty) is motivated by the data.
3. Generation is Harder than Classification: Generating faithful evidence imposes a burden beyond label prediction. Systems that produce correct verdicts often fail to provide the necessary evidentiary support, and conversely, fluent evidence can be generated for incorrect verdicts.
4. Evaluation Standards: The study advocates for evaluating retrieval as an intervention (via Bio-GRACE) rather than just a retrieval metric, and for maintaining transparency about failed generations and low-coverage outputs rather than silently discarding them.

The authors conclude that while fine-tuning currently offers the most reliable path for evidence-generating systems, future work must address the source-matching problem to make RAG effective for broader public health claims, potentially requiring human review when adequate evidence is unavailable.