Active Inference Guides AI Agents in Context Acquisition

Interactive AI agents often face the challenge of efficiently acquiring the right context when users omit crucial information like constraints or preferences. Agents must decide whether to proceed with default assumptions or expend resources on clarifying questions, retrieval calls, or tool invocations. Researchers propose framing this tradeoff as active inference for context acquisition. In this framework, an inner inference step continuously updates the agent's beliefs about a latent task state. An outer decision-making process then selects the most optimal next action—be it acquiring more context, performing a task action, or stopping—with the goal of minimizing expected free energy, while also considering token costs. In deterministic settings, the epistemic term of this framework simplifies to expected information gain, potentially normalized by token cost. The concept is instantiated in Optimal Question Asking (OQA), which uses exact posteriors and dynamic programming. Benchmarking frontier language models on various categorical tasks demonstrates the framework's effectiveness in guiding agents to efficiently gather necessary context, including clarification before generation and automated prompt optimization under budget constraints. This model-agnostic formulation positions active inference as a core design principle for AI agent context acquisition layers.