2651 - Show Your Work: Verbatim Evidence Requirements and Automated Assessment for Large Language Models in Biomedical Text Processing
Presenter(s)
P. Windisch1,2, J. Weyrich3, F. Dennstaedt2, D. R. Zwahlen4, R. Foerster1,2, and C. Schroeder1,2; 1Department of Radiation Oncology, Cantonal Hospital Winterthur, Winterthur, Switzerland, 2Department of Radiation Oncology, Inselspital, Bern University Hospital, University of Bern, Bern, Switzerland, 3Faculty of Medicine, University of Bern, Bern, Switzerland, 4Department of Radiation Oncology, Kantonsspital Winterthur, Winterthur, Switzerland
Purpose/Objective(s): Large language models (LLMs) are used for biomedical text processing, but individual decisions are often hard to audit. We evaluated whether enforcing a mechanically checkable “show your work” quote affects accuracy, stability, and verifiability for trial eligibility-scope classification from abstracts.
Materials/Methods: We used 200 oncology randomized controlled trials (2005 - 2023) and provided models with only the title and abstract. Trials were labeled with whether they allowed for the inclusion of patients with localized and/or metastatic disease. Three flagship models (GPT-5.2, Gemini 3 Flash, Claude Opus 4.5) were queried with default settings in two independent conditions: label-only and label plus a verbatim supporting quote. Models could abstain if they deemed the abstract to not contain sufficient information. Each condition was repeated three times per abstract. Quotes were mechanically validated as exact substrings after whitespace normalization, and a separate judge step used an LLM to rate whether each quote supported the assigned label.
Results: Evidence requirements modestly reduced coverage (GPT-5.2 86.2% to 84.3%, Gemini 98.3% to 92.8%, Claude 96.0% to 94.5%) by increasing abstentions and, for Gemini, invalid outputs. Conditional macro-F1 remained high but changed by model (slight gains for GPT-5.2 and Gemini, decrease for Claude). Labels were stable across repetitions (Fleiss’ kappa 0.829 to 0.969). Mechanically valid quotes occurred in 83.3% to 91.2% of runs, yet only 48.0% to 78.8% of evidence-bearing predictions were judged semantically supported. Restricting to supported predictions increased macro-F1 at the cost of lower coverage.
Conclusion: Substring-verifiable quotes provide an automated audit trail and enable selective, higher-trust automation when applying LLMs to biomedical text processing. However, this approach introduces new failure modes and trades coverage for verifiability in a model-dependent way.