Structured Behavioral Heterogeneity as Latent Regime Constraints

Sequential decision data often exhibit structured behavioral heterogeneity: similar observed conditions can lead to different actions across trajectories and time. Standard approaches typically attribute this variation to drifting rewards or unstructured noise, which can produce models that are difficult to interpret and compare across settings. We instead model this heterogeneity as arising from latent constraint regimes, which induce persistent, regime-dependent deviations from a shared decision anchor. We introduce a framework that decomposes behavior into (i) a shared anchor value function capturing stable, outcome-relevant decision logic and (ii) a switching regime process that generates sparse, interpretable deviations from this anchor. This decomposition separates stable decision structure from systematic variation in observed actions. To construct the anchor under partial observability and irregular decision timing, we use a continuous-time latent-state model that infers belief states from irregularly timed decision data and produces comparable action values across trajectories. The regime layer then captures structured deviations without modifying the underlying objective. Across clinical care and retail pricing datasets, the proposed approach improves held-out action prediction relative to stationary models and interpretable behavioral baselines, while identifying persistent regime assignments and sparse, feature-level deviations. These findings show that structured heterogeneity can be captured by a shared anchor with switching, regime-dependent deviations.