Multimodal Large Language Model Discrimination of Poster-to-Manuscript Conversion at the 2022-2023 Society of Surgical Oncology Annual Meetings.
Nov 11, 2026DOI: 10.1245/s10434-026-20905-9Journal ArticleASO
Written by: Gavin McAfeeJayati AtaharAmanda KaineOlzhas ZhorayevNaoru KoizumiMeng-Hao LiJorge OrtizSohail Khan
BACKGROUND: Whether a multimodal large language model (MLLM) can identify publishable scientific work is untested. This study evaluated whether GPT-4.1 mini could distinguish Society of Surgical Oncology (SSO) posters that reached full-text publication using poster content alone and examined characteristics associated with conversion.METHODS: This study retrospectively analyzed 597 posters from the 2022-2023 SSO Annual Meetings. Each poster was scored by GPT-4.1 on five research-quality domains during 30 runs at temperature zero. Discrimination was assessed by the area under the receiver operating characteristic curve (AUC). Multivariable logistic regression tested whether the score added information beyond poster and investigator characteristics.RESULTS: Of 597 posters, 295 (49.4%) reached full-text publication. The composite score showed modest discrimination (AUC, 0.62; 95% confidence interval [CI], 0.57-0.66), with no evidence that performance differed across categories (interaction P = 0.90). Each 5-point increase in score was associated with higher adjusted odds of publication (odds ratio [OR], 1.45; 95% CI, 1.21-1.75), and adding the score to poster characteristics improved discrimination (ΔAUC, 0.05; 95% CI, 0.01-0.08). Design rigor, data quality, conclusion validity, and publication potential scores were associated with publication, but novelty was not. Multicenter status, first-author training status, and author sex were not significantly associated with publication. Lower adjusted odds for Hispanic first authors did not persist after multiplicity correction (OR, 0.35; 95% CI, 0.15-0.81).CONCLUSIONS: The GPT-4.1 mini scores modestly distinguished posters that reached publication and added information beyond poster characteristics. Multimodal large language models merit further study as instruments for examining and strengthening scholarly work rather than as arbiters of publishability.