When a broad clue makes an AI guess

A question asks which disease a drug treats. The retrieved passage says only that it treats hyperlipidemia, a broader disease category. That is useful information: it narrows the possibilities. But it does not establish which specific condition is meant. Should an AI answer, decline, or consult a structured medical resource?

S M Rayeed, Charles V. Stewart and Yao Ma examine what happens when retrieval-augmented language models encounter this kind of incomplete clue. Retrieval-augmented generation, or RAG, combines a language model with text retrieved to help answer a question. The authors call a broad-but-relevant clue “resolvable insufficiency,” because an ontology—a structured map of categories and their members—could help identify candidate answers. They call it “Clade” when a model instead answers directly from the coarse clue.

Testing the effect of a broad clue

The researchers grouped passages into three types: those naming the exact answer, those naming only a broader category, and those mentioning neither. They tested five instruction-tuned open models on multiple-choice questions about drug–disease relationships in PrimeKG, a biomedical knowledge graph, using the MONDO disease ontology. Each question included an option for saying there was not enough information. They also tested questions about species interactions in GloBI and biological taxonomy.

To separate the effect of a broad clue from the difficulty of the questions, they compared each question with and without its passage, then compared those changes across passage types. They also edited passages for the same questions: replacing the specific answer with its broader category, or deleting the sentence. These comparisons help assess whether the category mention itself changes model behavior.

Fewer refusals, more wrong guesses

In the biomedical test, the three models in the main pooled analysis abstained less when given a family-only passage than when given a silent one. The difference was 13.8 percentage points. Compared with the silent-passage condition, the broad clue led to increases in both correct and wrong answers. The authors describe about half of the additional answers as wrong; that does not mean half of all answers were incorrect. Precision—the share of attempted answers that were right—was 45–59% on family-only passages, versus 84–95% on passages naming the answer.

The edited-passage test further supports the role of the broader term. For the same questions, keeping the family name rather than deleting the sentence reduced abstention by 17.4 points. In the ecological test, family-only passages led to less abstention than silent passages. A separate counterfactual edit found that keeping a genus name rather than deleting the sentence also reduced abstention. The comparisons are related, but they are not the same test. In a smaller free-text check, no model supplied the exact answer from a family-only passage; some gave the broad category as though it were the answer.

The study also tested possible safeguards. Confidence scores did a poor job separating correct from incorrect guesses on family-only passages. A larger version of one model did not eliminate the behavior. An instruction to abstain on broad categories increased abstention across passage types in the pooled results, but responses varied substantially by model: Granite changed little, and its abstention decreased on family-only and silent passages. The instruction also reduced correct answers on passages that named the answer. These findings show that the tested measures did not reliably solve the problem in these experiments, not that they can never help.

Ontologies help—but the hand-off is hard

An ontology can turn a category into a list of candidate diseases. When researchers supplied the correct family, the resulting list contained the sampled answer in 98.5% of cases. Automatically identifying the relevant family from a drug’s page was harder: simple string matching produced a usable list containing the answer in 19.8% of cases. A model choosing among families mentioned on the page raised that figure to 30.0%. End-to-end recovery rose from 10.0% to 26.8% in the reported comparison, but remained incomplete. Within a list built from the correct family, a simple popularity-based choice outperformed the tested model chooser.

The takeaway is not that AI should always refuse when evidence is incomplete. Rather, the kind of incompleteness matters: a broad category can point toward useful candidates, but it does not prove which specific answer is correct. A system would need to identify the relevant category, consult the ontology, and avoid choosing when the evidence does not distinguish among candidates. The experiments suggest that these steps remain difficult.

There are important limits. The tests used particular databases, text sources, models and question formats; the ontology-recovery experiment was limited to PrimeKG. Multiple-choice questions with an explicit abstention option may not capture every real-world interaction. The authors also note that graph-based scoring can mark a medically valid answer wrong if the graph does not record it. The results describe benchmark behavior, not clinical performance.

The supplied record identifies the work as an arXiv v1 preprint and includes a 2027 AAAI copyright notice. The provided materials do not establish whether the work has been peer reviewed or published in a journal, so its review status is unconfirmed. The findings should be treated as provisional and checked against the original source before being relied on.