AI Paper Atlas

Research map / Reinforcement learning / Policy gradient and actor-critic

Safety Generalization Under Distribution Shift in Safe Reinforcement Learning: A Diabetes Testbed

Minjae Kwon, Josephine Lamp, Lu Feng

ICML 2026 · Reinforcement learning › Policy gradient and actor-critic · approach: Safety layers and shielding

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
evaluate safe RL policies on unseen patients instead of the training patientsafe RL algorithms (CPO, CUP, FOCOPS, CRPO, PPO-Lag, RCPO, TRPO-Lag) · diabetes simulator OOD patientsTIR (%): CPO 87.28 -> 76.73; PPO-Lag 85.29 -> 75.20Hurt
evaluate PCPO on unseen patientsPCPO · diabetes simulatorTIR (%): 36.74 -> 38.75No clear effect
add test-time predictive shield (ablation)PPO-Lag · T2D with pumpTIR (%): +14.15Improved
add test-time predictive shield (ablation)safe RL algorithms (avg) · T1D with pumpTIR (%): +6.08 averageImproved
add test-time predictive shield (ablation)CUP and PCPO · T1D with pumpTIR (%): -0.22 (CUP), -0.14 (PCPO)No clear effect
add test-time predictive shield (ablation)safe RL algorithms · T1D with pumpCV (%): FOCOPS -4.26Improved
add test-time predictive shield (ablation)PPO-Lag, RCPO, TRPO-Lag · T2D without pumpCV (%): PPO-Lag +4.39, RCPO +4.95, TRPO-Lag +5.49Hurt
add test-time predictive shield (ablation)CRPO · T2D without pumpTIR (%): +11.80 (CV -7.28)Improved
add test-time predictive shield (ablation)CUP and CRPO · T2D with pumpCV (%): CUP +0.53, CRPO +0.04Hurt
use rule-based shield instead of predictive shield (ablation)safe RL policies · all cohortsTIR and Risk IndexHurt
use rule-based shield instead of predictive shield for poor base controllers (ablation)CUP, PCPO · all cohortsTIR and Risk IndexMixed
replace NODE or ITransformer with BA-NODE dynamics predictor (ablation)glucose predictor · 24-step horizonMAE: NODE 3.18 -> 2.82Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy