Generate, Verify and Select:Evolutionary Fitness Shaping for Safe Clinical Text Generation
1Interdisciplinary Advanced Research Centre for Transformative Technologies (IARCT2), Mahindra University.
2Department of Computer Science and Engineering, Mahindra University.
Overview
Generate–Verify–Select (GVS) is an inference-time framework for safe clinical text generation, designed for Brief Hospital Course and Discharge Instructions. Rather than trusting a single model output, GVS produces a diverse candidate pool and applies explicit safety checks for medication hallucinations, numeric inconsistencies, negation flips, and sentence-level contradictions.
Final selection is driven by evolutionary fitness shaping: a Differential Evolution-tuned linear fitness function balances quality and safety at inference time. On Discharge Me, GVS improves official scores to 0.323 for BHC and 0.378 for DI while reducing medication hallucinations to 9.2–11.1% and roughly halving numeric and negation errors, without retraining the base generators.
Code
Coming Soon
Repository and implementation details will be added here.
Paper Link
Coming Soon
The publication link will appear here once it is available.