PoS is an inference-time framework for LLM agents that replaces reliance on raw history or compressed memory with an explicit belief state: a structured estimate of the current world state plus the unresolved task requirements. At each step, PoS validates that the belief remains internally consistent and supported by evidence, and it monitors whether actions are making meaningful progress. If not, it detects a "Belief Trapping" pattern and triggers recovery tailored to the type of unresolved requirement.
Across four benchmarks spanning execution and diagnosis, PoS achieved the highest overall performance with all three LLM backbones. Ablations showed that consistency validation and recovery both matter, and context-scaling experiments indicated the approach remains effective as context grows. The work positions belief construction and maintenance, not history retention or compression, as the central problem for long-horizon agents.