Evidence Use and Identifier-Conditioned Prior Knowledge in Large Language Model Classification of Oncology Trials...

Large language models (LLMs) can accurately classify biomedical documents, but strong benchmark performance does not establish that predictions are grounded in the supplied text. In biomedical literature tasks, titles, abstracts, digital object identifiers (DOIs), journal metadata, and trial identifiers may have been seen during pretraining and can trigger parametric knowledge or learned associations.

This study aimed to test whether oncology randomized trial success classification is driven by abstract evidence or by identifier-conditioned prior knowledge, and assess whether models follow counterfactual outcome evidence when it conflicts with original trial identifiers.

We evaluated 250 two-arm oncology randomized controlled trials from 7 major journals published between 2005 and 2023, each with a single primary endpoint and previously adjudicated positive or negative ground-truth label. The corpus included 58.4% (146/250) positive and 41.6% (104/250) negative trials. GPT-5.2, Gemini 3 Flash, and Claude Opus 4.5 were queried via vendor APIs under default settings using a single-token output instruction. For each trial, we created 5 deterministic input conditions: title+abstract, title only, DOI only, counterfactual title+abstract in which the primary endpoint outcome statement was minimally flipped, and the same counterfactual input paired with the original DOI to create an identifier-text conflict. Performance was assessed using valid format rate, accuracy, sensitivity, specificity, and F1-score.

The models showed high format adherence, with valid prediction rates of 97.2% to 100%. In the title+abstract condition, all models achieved high and balanced performance (accuracy and F1-score=0.96-0.97; sensitivity=0.96-0.97; specificity=0.96-0.98). Removing evidence reduced performance stepwise: title-only accuracy and F1-score fell to 0.79 to 0.88, and DOI-only performance fell to 0.63-0.67, exceeding the 58.4% majority class baseline but indicating limited identifier-driven signal. Counterfactual edits were concentrated in outcome-bearing text, with the Results and Conclusions sections modified for all trials, whereas the titles and Methods sections required edits in only 5.2% (13/250) and 1.6% (4/250) of trials. Against inverted labels, models followed counterfactual evidence with near-ceiling performance (accuracy and F1-score=0.96-0.99). Reintroducing the original DOI caused little change for GPT-5.2 (accuracy and F1-score=0.99) but modestly reduced F1-scores for Gemini (0.97) and Claude (0.95), mainly through lower sensitivity.

The evaluated LLMs robustly followed explicit end point statements in abstracts, including when those statements contradicted original trial outcomes. However, above-chance title-only and DOI-only performance, together with small decrements under counterfactual DOI conflicts, showed that identifiers can carry predictive signal and occasionally compete with textual evidence. Progressive content removal combined with counterfactual identifier-text conflicts offers a practical, reproducible audit for grounding in biomedical LLM evaluations.