arxiv:2610.01415

Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States

Published on Oct 1

· Submitted by

Yu Luo on Oct 2

#3 Paper of the day

alibaba

Upvote

Authors:

Abstract

Large language model (LLM) agents can now undertake increasingly complex tasks, but the way they organize interaction history into memory does not ensure a coherent understanding of the current world. We introduce PoS, an inference-time framework that constructs and continually maintains explicit belief states as the agent's decision context. Each belief combines an estimate of the current world state with unresolved task requirements, making explicit what the agent still needs to learn and accomplish. To keep this belief reliable and actionable, PoS validates its consistency and monitors task progress to detect Belief Trapping, where the agent continues to act without making meaningful progress toward the goal. Recovery is then tailored to both the trapping pattern and the type of unresolved task requirement. Experiments on four benchmarks spanning execution and diagnosis show that PoS achieves the highest overall performance on every benchmark with all three LLM backbones. Ablations demonstrate the importance of consistency validation and recovery, while context-scaling experiments show resilience to context growth. Together, these results support belief construction and continual maintenance as a foundation for long-horizon context management beyond history retention and compression.

View arXiv page View PDF Project page GitHub 16 Add to collection

Community

IDKTreeDP

Paper submitter about 13 hours ago

Long-horizon agents need to keep track of the current world state to guide their decisions. Yet maintaining a coherent belief alone does not ensure progress: agents can keep acting without meaningfully advancing their goals, a failure mode this paper calls Belief Trapping. This paper introduces PoS (Progression of States), an inference-time framework that combines explicit belief maintenance with trapping detection, diagnosis, and recovery, without additional training. PoS represents the current world and unresolved information and task requirements, validates belief updates for consistency and evidential support, and monitors whether interactions advance the task. When progress stalls, it diagnoses both the trapping pattern—stagnation, cycles, or drift—and the blocked requirement, then composes targeted recovery constraints to guide subsequent actions. Across four benchmarks covering task execution and evidence-seeking diagnosis, PoS achieves the highest overall performance among the evaluated methods in all 12 benchmark–backbone combinations, with relative gains over the strongest same-backbone baseline reaching 22.68% on ALFWorld and 37.89% on RCA-100 joint accuracy. These results highlight the value of continually maintaining beliefs and restoring meaningful task progress for reliable long-horizon agents.