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

What it claims

PoS (Progression of States) is a training-free, inference-time wrapper for LLM agents. It replaces raw interaction history with a maintained belief state, and adds a monitor that detects "Belief Trapping" and steers the agent out of it. The two headline figures are relative gains over the strongest same-backbone baseline, not absolute accuracy points:

• 22.68% on ALFWorld.
• 37.89% on RCA-100, where the metric is joint accuracy.

PoS is reported as best overall in all 12 benchmark–backbone combinations (4 benchmarks × 3 LLM backbones).

How it works

The page adds three mechanisms the summary only names:

• Two kinds of unresolved requirement. The belief tracks the current world state plus unresolved information requirements and unresolved task requirements. In other words, it tracks both what the agent still needs to learn and what it still needs to do.
• Validated updates. Each belief update is checked for consistency and for evidential support before it is accepted. It is not simply appended.
• Two-axis recovery. When progress stalls, PoS diagnoses two things: the trapping pattern (stagnation, cycle or drift) and which requirement is blocked. It then composes recovery constraints from both, and those constraints restrict the agent's next actions.

flowchart TD
    A[Agent action] --> O[Observation]
    O --> U[Propose belief update]
    U --> V{Consistent and supported?}
    V -- no --> U
    V -- yes --> B[Belief: world state + unresolved requirements]
    B --> P{Making progress?}
    P -- yes --> A
    P -- no --> D[Diagnose pattern + blocked requirement]
    D --> R[Recovery constraints]
    R --> A

Evidence

• Benchmarks. The four benchmarks fall into two families: task execution (e.g. ALFWorld) and evidence-seeking diagnosis (e.g. RCA-100). The page does not name the other two.
• Ablations. Removing consistency validation or removing recovery hurts performance. The page gives no magnitudes for either.
• Context scaling. Experiments report that PoS is resilient as context grows. This is the claim most relevant to anyone replacing history compression or summarisation.

Limits

• Abstract only. This page carries the abstract and an author comment. It does not give:
• the identities of the three backbones;
• the names of two of the four benchmarks or their baselines;
• absolute scores;
• ablation or context-scaling numbers;
• any cost or latency overhead.

Validation and monitoring presumably add LLM calls, but the page does not quantify this.

• Relative gains only. The 22.68% and 37.89% figures are relative improvements. A large relative gain on a low baseline can mean a small absolute change, so check the full paper's tables before relying on them.
• No artifacts linked. No code, model, dataset or Space is linked from the page. An implementation would have to be built from the paper itself.