Answers are grounded in this paper and cite their sources.
This framework analyzes prediction errors and adapts monitoring in real-time, enhancing model reliability during shifts.
Two critical questions arise when monitoring world models under distribution shift: how long should we trust the current model (temporal validity), and what type of failure occurred (diagnostic classification)? We address both through a unified framework combining Adaptive Receding Horizons (ARH) with Dual-Horizon sign-structure analysis (DH). ARH dynamically adjusts the prediction horizon based on model reliability — extending it during stable periods and contracting it when errors accumulate, following a physically-motivated law of adaptive trust. DH analyzes the sign structure of prediction errors across short and long horizons to classify failure modes: noise bursts (symmetric, transient), gradual drift (asymmetric, progressive), and compound failures (mixed signatures). Experiments across synthetic and real-world time series demonstrate that ARH maintains prediction quality under non-stationary conditions while DH achieves robust failure classification. The sign-based diagnostic is magnitude-invariant, distinguishing failure types regardless of signal amplitude. The framework provides interpretable, real-time monitoring without requiring retraining or access to ground-truth failure labels. This archive contains the full LaTeX source, compiled PDF (24 pages), and complete reproducibility code including TCPD datasets and all experiment scripts.
No takes yet. Share an insight, caveat, or question.
Régis RIGAUD (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: