Multi-turn agents must decide what information to seek, how to use new evidence, and when they are certain enough to act. The authors formalize this as an Epistemic Decision Process, where actions produce observations that update the agent's posterior over a latent task variable. Good policies, they argue, choose actions that are useful under the current belief rather than merely correlated with eventual success.
To train such policies, they introduce ECHO, a clipped policy-gradient objective that assigns turn-level credit using posterior-sensitive rewards. In the Clue Selector Game and the CRAFT multiagent coordination game, ECHO substantially improves resolution, information gain, and efficiency compared with trajectory-level GRPO, while matching or exceeding frontier baselines on epistemic metrics such as grounding, recovery, and calibration.
This is a single-source report, so there are no independent results to compare. The paper is available on arXiv as 2606.29745.