Active inference with reusable state-dependent value profiles

Abstract

Optimal behavior in volatile environments requires agents to deploy different value-control regimes across hidden latent structures. However, representing independent preferences, biases, and action confidence for every situation is computationally and statistically intractable. We introduce value profiles: a compact set of reusable parameter bundles-comprising outcome preferences, policy priors, and action precision-assigned to hidden states in a generative model. As posterior beliefs evolve trial-by-trial, effective control parameters emerge through belief-weighted mixing. This enables state-conditional strategy recruitment without the need for independent parameterization of every context. We evaluate this framework in a probabilistic reversal learning setting using simulation-based model recovery and parameter recovery, comparing profile-based models against static and uncertainty-coupled precision baselines. Model comparison favors the profile-based architecture, with consistent parameter recovery demonstrating its structural identifiability. Beyond predictive fit, we show how the framework allows for a mechanistic attribution of behavior to specific adaptive channels, distinguishing between changes in what an agent prefers and how decisively it acts. Overall, reusable value profiles provide a tractable computational account of belief-conditioned control, offering a mode-like representational scheme for behavioral flexibility that is both identifiable and theoretically grounded.