Hiding the raw action log from the agent and showing only a Bayesian posterior beats ReAct, QMDP and POMCP.

arXiv 2609.10036 traces premature commitment, collapse onto the wrong hypothesis after one observation, and policy drift as history grows to one structural cause: the agent is a history-conditioned policy with no explicit belief over hidden state. Their Belief-State Engine sits outside the LLM, maintains a posterior over a POMDP's latent states, and exposes only that. The soundness proof requires that the raw history is never shown, which is an unusual and testable design constraint.