Researchers propose active inference framework to optimize context acquisition for AI agents

Researchers propose framing context acquisition for interactive AI agents as an active inference problem, where an agent must decide whether to proceed with a default assumption or spend tokens on a clarifying action such as a question, retrieval call, tool call, or prompt trial.

The framework separates an inner inference step that updates beliefs over a latent task state from an outer decision step that selects the next context action, task action, or stop action to minimize expected free energy under cost.

In deterministic settings, the epistemic term reduces to expected information gain, optionally normalized by token cost, providing a tractable objective for balancing information gain against computational expense.

The authors instantiate the framework in Optimal Question Asking (OQA) with exact posteriors and a dynamic programming oracle, and benchmark frontier language models on binary and multiway categorical tasks ranging from 25 to 300 candidates.

The study also examines clarification before generation and automated prompt optimization under explicit token budgets, highlighting the trade-offs between exploration (seeking missing context) and exploitation (proceeding with current information).

The formulation is explicitly model-agnostic and is positioned as a design principle for the 'context-acquisition layer' of AI agents, suggesting a generalizable approach to improving agent reliability and efficiency.