Daily Hypernovelty Lead · October 25, 2026
A Real Citation Can Still Point to the Wrong Evidence
A version-one preprint submitted to arXiv on October 21 asks a narrow but consequential question: when an AI system cites a PubMed identifier to support a biomedical claim, how often does that identifier point to a paper that is actually relevant? Mohammed Saqlain’s “Real PMIDs, Wrong Papers” tested two language models on 100 claims from the SciFact development set. Its central warning is straightforward. A citation can be real, resolvable and almost useless as evidence for the sentence beside it.

The pilot compared four ways of connecting two efficient Gemini models to medical literature: no retrieval, one PubMed search formed from the claim, an agent allowed to write and run up to four searches, and an oracle condition supplied with the benchmark’s gold abstracts. The models stayed fixed while the evidence available to them changed.
Verification bottleneck
In the closed-book condition, the models produced 335 PubMed IDs. Of the 327 IDs that existed, a separate language-model judge rated only three as relevant to the claims they were supposed to support. None matched the corresponding SciFact gold paper. The identifier passed an existence check while the evidence failed a relevance check.
One search did not close the gap. Among the 64 claims used for gold-recall analysis, single-shot retrieval found the gold abstract for four claims, or 6%, with either model. Accuracy reached 53% and 49%, compared with 55% and 59% when the models answered from their internal knowledge. Those differences were not statistically significant in this sample.
The agentic search harness did better at retrieval, finding the gold abstract for 19% to 20% of the claims, and both models reached 62% accuracy. Still, the differences among the realistic harnesses were not statistically significant at this sample size. The clearest separation appeared when the correct evidence was supplied. In the oracle condition, accuracy rose to 86% and 79%, gains of 24 and 17 percentage points over agentic search. The paper reports those differences as statistically significant.
The agents also left much of their search allowance unused. They averaged 1.42 and 1.70 searches per claim. Depending on the model, 71% or 55% of claims were answered after one search, while the full four-search budget was used in only 3% or 6% of cases. More searching would not automatically produce the right paper, but the pattern shows that granting a tool and setting a budget do not ensure that an agent will investigate deeply.
Citation relevance improved when citations were tied to actual retrieval. Across the agentic runs, the relevance judge rated about 67% of citations as relevant. Out-of-retrieval citations were 0% and 2.6% across the two models. That is a large improvement over real-looking identifiers generated from memory, though it still leaves a meaningful fraction of citations that need checking.
A useful verification receipt therefore needs more than an identifier. It should connect the claim, the search query, the retrieved paper, the supporting passage and the final verdict. That trace lets a reviewer inspect where the chain broke instead of treating a link to PubMed as proof that the evidence was read and applied correctly.
This remains a small benchmark study rather than evidence from clinical deployment. It used 100 claims, producing wide confidence intervals, two efficient models from one provider and one cross-family language model as the relevance judge. The authors also note that live PubMed can contain newer evidence than SciFact’s older gold set, and that title matching gives only a lower bound on recall. In a 50-pair check, the judge rated 21 of 25 gold pairs as relevant and none of 25 deliberately mismatched pairs as relevant. Its ratings should still be read as measured estimates rather than ground truth.
Opportunities
Publishers and research teams can require an evidence trace for AI-assisted claims: what was asked, what was retrieved, which passage supports the statement and who accepted the verdict.
Product builders can separate citation resolution from citation relevance. A green check for “this PMID exists” should never be displayed as if it also means “this paper supports the claim.”
Institutions testing evidence tools can measure an oracle gap of their own. Comparing performance with correct documents supplied against performance when the system must retrieve them reveals how much of the failure sits in search rather than final reasoning.
Reviewers and editors can sample the chain, not merely the bibliography. A small audit of claim-to-passage alignment may expose confident citation failures that reference-format checks will miss.