The abstract is read at the publisher; the summary is JClub's.
Unanimous selection by four predictor models increased the match rate of hypotheses generated by Grok 4.5 and Claude Opus 5.5 to withheld findings in 50 ageing biology studies, reaching 73.7% and 63.6% respectively.
- Why it matters: It tests whether large language models can generate and then identify hypotheses that align with published ageing biology findings, addressing the gap between plausible text and verifiable, literature-consistent claims.
- What they did: For 50 peer-reviewed ageing biology studies, four LLMs generated 2997 hypotheses using three prompting methods; hypotheses were matched to withheld reference findings by judge-model scoring, then filtered by a unanimous predictor rule.
- The result: Single-request match rates were 13.2%–18.0%, while Omega Point was 2.8%–7.6%; among unanimously selected hypotheses, match rates were 73.7% (Grok 4.5), 47.4% (Gemini 4 Argon), and 63.6% (Claude Opus 5.5), with pooled selected hypotheses exceeding first-listed by 23.2 pp (95% CI [13.0, 32.6]).
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.