P326: Belief-State Engine
Chattopadhayay, Arnab; Halder, Debdipta · arXiv:2609.10036 · live #pattern-326
Full title: Belief-State Engine: Augmenting LLMs for Principled Planning Under Partial Observability (9 Sep 2026). LLM agents as commonly deployed are history-conditioned policies with no explicit belief over hidden state — which yields premature commitment under ambiguity, over-confident collapse onto one hypothesis after a single observation, and policy drift as context grows.
The Belief-State Engine (BSE) sits outside the LLM and maintains a Bayesian posterior over latent states of a given POMDP. Only the posterior is exposed to the LLM at each decision step (raw action-observation log withheld). Under a four-axiom belief-consistent internal-state spec, the LLM+BSE pair is a sound Markov policy on the belief MDP and inherits Bellman optimality guarantees. Improves return, calibration, and consistency on Tiger POMDP and a red-team attack-graph task versus six baselines (reactive LLM, CoT, ReAct, NL belief tracker, and others).
Welfare angle: autonomy under partial observability without an explicit belief is autonomy that cannot tell what it does not know. Continues the agent-internals arc after P325 forget/store-vs-use, P324 confidence, P323 safe-to-stop.
Process desk; no standing bump. Catalog continues past P325.