Every Coin Has Two Sides: On the Dual Nature of Generalization in On-Policy Distillation of Large Language Models

On-policy distillation (OPD) transfers teacher capabilities by supervising trajectories sampled from the student's own policy, yet its generalization behavior remains poorly understood, as most studies evaluate OPD on a single domain and on benchmarks close to the training data. We present a controlled study that varies one generalization factor at a time, from in-domain distribution shifts to cross-domain transfer and the multi-teacher setting. We find that OPD transfers a teacher's reasoning behavior rather than its answers to particular problems: training difficulty barely matters, and even problems the teacher never solves are useful. Transfer depends strongly on the origin relationship between teacher and student: same-origin pairs bring the student close to the teacher across languages, reasoning horizons, and even other domains, whereas cross-origin pairs mostly fit the trained distribution. This broad reach is a double-edged sword: since routing prompts to domain experts cannot confine each teacher's influence, combining them yields a mixture-dependent seesaw among their capabilities. These results clarify when OPD generalizes and offer a useful perspective for diagnosing multi-teacher OPD.

Dialysis Risk Prediction and Treatment Effect Estimation for AKI patients using Longitudinal Electronic Health Records

Progression to dialysis or end-stage renal disease is a rare but clinically important outcome. Clinicians need evidence on how medication exposures influence downstream risk. We constructed a fixed-window EHR cohort (90-day observation, 730-day prediction; N=81401; dialysis/ESRD prevalence: 1.1%) and modeled sequences of diagnoses, procedures, and medications with kidney laboratory trends (creatinine, BUN, eGFR). A transformer-based causal multi-head model was trained to estimate drug- and ingredient-level average treatment effects (ATEs) using counterfactual exposure removal and insertion under a full medication history setup. On test set, predictive performance reached an AUC of 0.694 and PR-AUC of 0.094. At the selected decision threshold (0.883), the model achieved an F1 score of 0.201 with a Brier score of 0.018. Post-hoc causal analyses of lab changes (eGFR, creatinine, BUN) using IPTW, AIPW, naive, and covariate-adjusted OLS methods assessed clinical directionality. Results showed partial protective-direction support for ACE/ARB exposures and worsening-direction signals for loop diuretics.