Preprint reports comparative LLM method for cancer survival ranking

A preprint introduces CACSurv, a large-language-model framework that ranks cancer patients' relative survival prospects from clinical reports rather than predicting each patient's survival time independently. The authors say the method uses comparisons among patients and rewards based on relationships that remain valid when outcomes are censored. On their new TCGA-SurvReport benchmark, spanning six TCGA cancer cohorts, they report the highest C-index in every cohort and an average score of 0.722. The reported average exceeds the strongest published survival model by 6.5 percentage points.