Active Inference as Context Acquisition for AI Agents

From Suvrit Sra's group at TUM with UC Davis. Frames the context-acquisition layer of an agent harness — whether to act on the current context or spend tokens on a clarifying question, retrieval, tool call or prompt trial — as active inference: a bilevel one-step problem in which an inner step updates beliefs over a latent task state under a hypothetical observation and an outer step picks the next context action, task action or stop action to minimize expected free energy under token cost. In deterministic settings the epistemic term reduces to expected information gain, optionally normalized by cost, and value-of-information conditions give explicit stopping rules.

The framework is instantiated in Optimal Question Asking (OQA), an attribute-table benchmark with exact posteriors and a dynamic-programming oracle that measures a model's "planning gap" (extra questions relative to the oracle). Seven frontier models (GPT-5, GPT-4.1, Gemini 2.5 Pro, Gemini 2.0 Flash, Claude Sonnet 4.5, Claude Haiku 4.5, Grok 4) reduce uncertainty every turn on binary (25/100 candidates) and multiway (100–300 candidates) tasks but consistently ask more questions than the oracle. Two token-budget case studies — clarification before generation on 48 synthetic product tasks and automated prompt optimization — show that small amounts of targeted clarification or experimentation improve final success when they buy decision-relevant information. Submitted June 2026; appeared on arXiv August 2026.