Pragmatic Curiosity: Unifying Bayesian Optimization and Experimental Design via Active Inference

ICLR 2026 Conference Submission20460 Authors

Keywords: Bayesian Optimization, Bayesian Experimental Design, Active Inference

TL;DR: We introduce "pragmatic curiosity", a new paradigm that resolves the classic explore-exploit dilemma by treating goal-seeking and information-seeking as two facets of a single, principled objective.

Abstract: Bayesian Optimization (BO) and Bayesian Experimental Design (BED) have traditionally offered separate solutions for goal-oriented and information-oriented tasks, respectively, leaving a gap in complex problems where learning and optimization are not separate phases but deeply intertwined objectives. In this paper, we provide the first unified framework of BO and BED, which is rooted in the principles of active inference (AIF). We introduce "pragmatic curiosity", a new paradigm where the classic explore-exploit dilemma is resolved by minimizing a single objective: the Expected Free Energy (EFE), which naturally balances pragmatic (goal-seeking) and epistemic (information-seeking) drives. We demonstrate the power of this approach on a suite of challenging hybrid tasks, including constrained system identification, targeted active search, and composite optimization with unknown preferences. Empirical results prove the cross-domain adaptability and effectiveness of our proposed framework: our "pragmatic curiosity" paradigm consistently outperforms standard baselines in BO and BED, demonstrating quantifiable improvements in key metrics like estimation accuracy, critical region coverage, and final solution quality.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 20460