A new paper introduces PoS, an inference-time framework that constructs and continually maintains explicit belief states as an LLM agent’s decision context, rather than relying on raw interaction history. The framework validates consistency and detects “Belief Trapping,” situations where an agent acts without meaningful progress, triggering tailored recovery strategies. Across four benchmarks, agents using PoS outperformed existing context-management approaches, with ablations confirming that both the consistency validation and recovery mechanisms are essential to the gains.
