Qingfang Liu, Qiao Jin, Joe D. Menke +2cs.IR cs.AI cs.CL
Large language model (LLM) chatbots are increasingly used to answer clinical questions with citations to relevant clinical studies. Prior research has largely focused on citation fabrication, leaving a gap in evaluating the quality of retrieved studies and the factors driving their selection. In this study, we evaluated three general-purpose LLM chatbots: Claude Sonnet 5, Gemini 3.1 Pro, and ChatGPT GPT-5.5. We prompted the models with clinical questions adapted from 20 review questions in Issues 6 and 7 of the 2026 Cochrane Database of Systematic Reviews, simulating patient, clinician, and evidence-synthesis researcher roles. Each chatbot was queried under each user role with four independent repetitions, yielding 720 responses. Each chatbot was asked to support its answers with primary clinical citations, which we benchmarked against the included and excluded study sets of the Cochrane reviews. On average, a chatbot response retrieved 39.2% $\pm$ 29.8% of Cochrane included studies, while citing 5.0% $\pm$ 9.4% of excluded studies. Recall of Cochrane included studies varied significantly by model and user role. ChatGPT achieved higher recall than Claude or Gemini (63.1% $\pm$ 29.5% vs. 37.0% $\pm$ 23.8% vs. 17.3% $\pm$ 13.1%; $p=2.0\times10^{-5}$). The researcher role yielded higher recall than the clinician or patient roles (42.8% $\pm$ 30.8% vs. 38.6% $\pm$ 28.9% vs. 36.1% $\pm$ 29.3%; $p=2.0\times10^{-5}$). Controlling for publication year, citations per year, and open-access status, sample size was the only independently significant predictor of retrieval (odds ratio 1.80 per 1-unit increase in log sample size, 95% CI 1.37-2.36, $p=2.34\times10^{-5}$). These findings suggest that while LLM chatbots can retrieve some studies identified by expert reviewers, their performance varies by model and user role, and they exhibit a bias toward clinical trials with larger sample sizes.
Reported verdicts on GraphRAG versus vector RAG disagree, and the evidence is typically tied to a single corpus, embedder, and judge -- and, we show, to where citation quality is measured. We present a triple-robustness analysis that holds a five-pipeline architecture matrix fixed and varies embedder (local e5-small vs. Azure text-embedding-3-small), corpus (DO-178C typed-edge requirements vs. Wikipedia paragraph chains via MuSiQue), and judge (paired GPT-5.4 x GPT-4.1 on both corpora), over 2x4,440 main-matrix runs, 600 cross-corpus runs, and over 5,000 faithfulness judgments. (C2a) GraphRAG's graph walk floods the context window at precision 0.12-0.23, but the synthesizer cites selectively at precision 0.48-0.65; scoring the retrieved set as the attribution set inverts the architecture ranking, which reconciles part of the disagreement in prior reports. (C1) Answer-level citation winners are corpus- and stratum-conditional but embedder-robust: GraphRAG ties vanilla on short-hop DO-178C queries and wins every MuSiQue stratum, while agentic pipelines lead only on 3+-hop requirements queries. (C2b) Faithfulness is corpus-conditional: on DO-178C it declines with hop distance (trend p<0.05 in three of four judge x embedder combinations); on Wikipedia chains neither judge shows a collapse. (C3) Single-judge LLM faithfulness is fragile to retrieval state: GPT-5.4's self-kappa across embedders is 0.137 (41% verdict change) against a same-day test-retest floor of 0.76, and re-judging frozen inputs eleven weeks later gives kappa <= 0.14 for both judges. A learned router on dense embeddings alone reaches macro-F1 0.86 on hop classification (C4). We argue that RAG architecture claims should be tested at this level of robustness -- including robustness to the citation-measurement point -- before they are trusted.
Retrieval-augmented generation systems for legal question answering typically retrieve passages based on semantic similarity and provide them to a language model, which then generates cited answers. Prior work assumes that highly ranked passages are most likely to be usefully cited by the model. Perturbation-based attribution methods, such as C-LIME, have been used exclusively for post-hoc explanation. However, on the AQuAECHR benchmark, semantic similarity does not correlate with passage attribution. Within a retriever's candidate pool, similarity-based ranking performs worse than random selection at surfacing gold citation paragraphs. To address this limitation, a lightweight cross-encoder is trained on continuous perturbation-based attribution scores to re-rank passages prior to generation. This approach is evaluated on the AQuAECHR benchmark, using two language models and five-fold cross-validation. The re-ranker substantially improves citation faithfulness and alignment with gold expert answers. Notably, two re-rankers trained independently on different models converge beyond their raw attribution agreement. This finding indicates that the cross-encoder reduces model-specific noise and produces a shared relevance signal that partially transfers across models, although same-model re-ranking remains more effective. These results demonstrate that perturbation-based attribution provides a practical, model-agnostic training signal for citation-aware retrieval.