Beyond Memory: PoS Gives Long-Horizon Agents Explicit, Checkable Beliefs

Introduction

For a long-horizon agent, the central challenge is not simply remembering what happened. As an interaction grows, the agent must keep track of the current state of the world, distinguish known facts from unresolved information, and determine whether each action actually moves the task forward. Retaining or compressing the conversation history alone does not guarantee such a coherent, actionable understanding.

The paper introduces PoS, or Progression of States, an inference-time framework that treats context management as continual belief maintenance rather than history storage. Its main representation is an explicit belief state that guides the agent’s next decision.

How PoS works

• A joint view of the world and the task. Each belief contains an estimate of the current world, unresolved information, and outstanding task requirements. The agent can therefore represent not only what is happening, but also what still needs to be learned or accomplished.
• Consistency and evidence checks. After interaction with the environment, PoS validates whether the updated belief remains internally consistent and supported by available evidence. This is intended to prevent unsupported inferences from becoming the agent’s working reality.
• Progress monitoring. An agent may continue taking plausible actions without narrowing uncertainty or advancing toward the goal. PoS names this failure mode Belief Trapping.
• Pattern-specific recovery. The framework distinguishes trapping patterns such as stagnation, cycles, and drift. It also identifies the type of unresolved requirement that is blocking progress, then composes recovery constraints tailored to that situation.

Results and broader significance

The evaluation covers four benchmarks spanning task execution and evidence-seeking diagnosis, with three LLM backbones. PoS achieved the highest overall performance in all 12 benchmark-backbone combinations. Relative to the strongest baseline using the same backbone, the reported gains reached 22.68% on ALFWorld and 37.89% in RCA-100 joint accuracy. Ablations indicate that both consistency validation and recovery matter, while context-scaling experiments suggest that the framework remains resilient as the interaction context grows.

The broader contribution is a shift in how agent memory is framed. The goal is not to retain as much history as possible, but to maintain a current belief that can be checked, revised, and used for decisions. This is especially relevant to systems that plan over many steps, interact with environments, or gather evidence for diagnosis. At the same time, the supplied material primarily reports benchmark results. How much belief quality depends on the underlying model’s reasoning ability, and how PoS behaves in more open-ended environments, remain important questions for future work.

Source: Hugging Face Daily Papers