When Loss Signals Dominate Context: Adaptive Expert Routing in the Loss-Dominance Regime

Vatsal Khanna, Samridhi Saini

Keywords: Adaptive Expert Routing, Online Learning, Full Information Feedback, Contextual Bandit, Information Rate Mismatch

TL;DR: When feedback is fast and expert differences are large, loss signals alone are sufficient and learned context can hurt.

Abstract: Modern adaptive systems increasingly rely on learned contextual models, often deep networks to anticipate regime changes and guide decision-making. Intuitively, such context should improve performance by accelerating the identification of the best expert. However, this intuition implicitly assumes that learned signals provide information faster or more reliably than the system’s own feedback. In this work, we show that this assumption fails in a natural and identifiable class of settings. We study adaptive expert routing under full-information feedback and identify a regime in which online loss signals alone are sufficient for near-optimal performance. When inter-expert performance gaps are large, loss-based routing rapidly concentrates on the optimal expert, leaving little room for improvement from contextual predictions. In this regime, even moderately accurate predictors introduce persistent bias that outweighs their potential benefit. We formalize this phenomenon through a mixing-time analysis and a regret decomposition, leading to a computable condition under which contextual information does not improve routing performance in this regime. This characterization yields a practical diagnostic based on context-weight interpolation that determines whether a given system operates in this loss-sufficiency regime. Empirical evaluation in a real-time, microsecond-latency adaptive control system (16 kHz Active Noise Control) confirms the theory: loss-based routing matches oracle performance, while contextual routing does not provide measurable improvement and often degrades performance in our setting.

Submission Number: 81