Stable Multi-Step Rollouts via Uncertainty-Guided Hybrid Dynamics

What the math gives to ML

The paper offers a constructive hybrid-rollout mechanism: use a provably contracting nominal predictor inside a calibrated uncertainty region, and activate a flexible excursion predictor only when epistemic uncertainty exceeds a threshold. The transferable asset is the separation of stability regimes: contraction limits recursive error accumulation in familiar states, while the excursion branch is restricted to bounded or affine growth outside that region. This can be implemented in recurrent sequence models, state-space models, and model-based RL world models by spectrally constraining the nominal branch and using an uncertainty-calibrated gate during multi-step prediction. The main falsifiable prediction is a change in rollout-growth behavior near the uncertainty threshold, with bounded error below it and slower divergence than an ungated flexible model above it.