CalTwin: A Fisher information regularization method for calibrated and distribution-shift-robust medical world models.

Khan et al. propose CalTwin, a lightweight regularization objective for GRU-based medical world models that addresses both covariate shift and confidence miscalibration. On PhysioNet 2019 Sepsis, CalTwin reduces OOD next-step hidden state MSE by 9.1%, with the Fisher information matrix penalty contributing 7.0% of the reduction, but CMP alone yields a larger ECE reduction (1.3% vs. 0.7%). On eICU-CRD, the FIM penalty performs 0.4% worse than the baseline, and CMP alone outperforms CalTwin on four of six primary metrics.