Calibration-risk routing for controlled world-model adaptation
MC-WM partitions target data to select lower-calibration-risk world models and weights imagined policy updates via learned confidence, evaluated across 541 MuJoCo shift executions.
Published Oct 1, 2026arXiv ↗
Only vote on papers you've read. Sign in with GitHub to vote.
Calibration-risk routing cleanly partitions fit, selection, and calibration to choose lower-risk world-model families with elegant validity predicates, though 541 runs may obscure a brittle routing rule hidden by one repeated artifact-gate cell.
Abstract
Model-based reinforcement learning (MBRL) can exploit simulated experience, but a simulator-to-target shift creates a model-selection problem: correcting the simulator and fitting the target directly can each fail under limited target data. We introduce the Model-Corrected World Model (MC-WM), which separates initial target data into disjoint fit, selection, and calibration partitions and deploys the family with lower standardized calibration risk. A learned confidence signal and deterministic validity predicates weight one-step imagined policy updates without rewriting physical rewards. We evaluate 540 unique reported run cells across three controlled Multi-Joint dynamics with Contact (MuJoCo) shifts; one exact-routing cell was repeated after a pre-deployment artifact gate, giving 541 completed executions.