Hide the Action Log From the Agent and Show It Only a Bayesian Posterior

arXiv 2609.10036·medium signalXBlueskyLinkedInCopy link

arXiv 2609.10036 (2026-09-09) traces characteristic LLM agent failures under partial observability (premature commitment on ambiguous feedback, collapsing onto the wrong hypothesis after one observation, policy drift as history grows) to a single structural cause: the agent is a history-conditioned policy with no explicit belief over hidden state. The Belief-State Engine sits outside the LLM, maintains a Bayesian posterior over a POMDP's latent states, and exposes only that posterior while withholding the raw action-observation log. The authors prove soundness as a Markov policy on the induced belief MDP provided the raw history is never shown, and beat six baselines including ReAct, QMDP and POMCP on the Tiger POMDP and a red-team attack-graph task.