DyBBT: Dynamic Balance via Bandit-inspired Targeting for Dialogue Policy with Cognitive Dual-Systems

ICLR 2026 Conference Submission15050 Authors

Keywords: dialog policy, dual-systems, task-oriented dialog, exploration strategy

TL;DR: DyBBT is a dialog policy framework that leverages a cognitive dual-system architecture and a bandit-inspired meta-controller to dynamically balance exploration and exploitation, achieving SOTA performance across multiple benchmarks.

Abstract: Task oriented dialog systems often rely on static exploration strategies that do not adapt to dynamic dialog contexts, leading to inefficient exploration and suboptimal performance. We propose DyBBT, a novel dialog policy learning framework that formalizes the exploration challenge through a structured cognitive state space $\mathcal{C}$ that captures dialog progression, user uncertainty, and slot dependency. DyBBT proposes a bandit inspired meta-controller that dynamically switches between a fast intuitive inference (System 1) and a slow deliberative reasoner (System 2) based on real-time cognitive states and visitation counts. Extensive experiments on single- and multi-domain benchmarks show that DyBBT achieves state-of-the-art performance in success rate, efficiency, and generalization, with human evaluations confirming that its decisions are well aligned with expert judgment. The code is available at \href{https://anonymous.4open.science/r/DyBBT-C6B7}{https://anonymous.4open.science/r/DyBBT-C6B7}.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 15050