Back to feed
News Story
机器之心
1 sources

Nankai University et al. Propose PoS to Address Belief Trapping in Agents

Researchers from Nankai University, Alibaba Cloud, and Tsinghua University have proposed PoS (Progression of States) to address Belief Trapping in long-horizon agents, where agents keep acting but key requirements fail to progress. PoS maintains explicit belief states, measures progress via Belief Health, and applies Trapping-Aware Recovery when stagnation is detected. Across 12 experimental settings on four benchmarks and three base models, PoS outperformed compared methods on overall metrics.

SynthePulse Insight · AI deep readingMembers

When an Agent Keeps Acting but Gets Nowhere: PoS Tries to Make "Stuck" a Computable Problem

Version 1 · 1 source

Nankai University, together with Alibaba Cloud and Tsinghua University, proposes PoS, breaking down long-horizon agents' Belief Trapping from a vague feeling of failure into three operable parts: explicit belief states, belief health, and trapping-aware recovery. The paper self-reports the highest overall metrics within the comparison scope across 12 settings spanning four benchmarks and three base models, but the costs and external validation remain unresolved questions.

  • The research team defines Belief Trapping as: the agent keeps interacting with the environment, but the goal-relevant evolution of its beliefs can no longer produce meaningful task progress.
  • PoS consists of three parts: continuously maintaining explicit belief states, using Belief Health to measure progress, and applying Trapping-Aware Recovery once trapping is detected.
  • In the heated-cup case, the Raw Trajectory baseline still failed after using up 50 steps, while PoS completed it within 17 steps; the baseline had repeated the ineffective cabinet-checking behavior 36 times.
Open section navigationA Cup That Missed "Heating" Exposes the Blind Spot of Long-Horizon Agents

A Cup That Missed "Heating" Exposes the Blind Spot of Long-Horizon Agents

In the ALFWorld heated-cup task shown in the paper, the agent first puts the cup into the microwave, then into the cabinet, and then falls into a loop of repeatedly checking the cabinet. This ineffective behavior repeats 36 times until the action budget is exhausted, and the task is still incomplete—what was missed is precisely the "heating" step. The Raw Trajectory baseline, which uses the full interaction history, still fails after using up 50 steps, whereas PoS completes the same task within 17 steps.

The value of this case is not the cup itself, but that it puts a class of failure modes on the table: the agent is still calling tools, receiving observations, and generating next-step plans, appearing to be working the whole time, yet the key requirement is never advanced. The research team explicitly names this phenomenon Belief Trapping: the agent keeps interacting with the environment, but the goal-relevant evolution of its beliefs can no longer produce meaningful task progress. It can manifest as repeating in place, cycling among a few states, or continuously exploring without ever touching the truly unresolved problem.

What needs to be distinguished is this: action repetition itself does not equal being trapped. Multiple calls to the same tool may be collecting different evidence; conversely, switching tools at every step may also keep avoiding the key problem. This is exactly the judgment difficulty that PoS aims to address.

PoS's Three-Part Structure: Write It Clearly First, Then Judge, Finally Intervene

The first part of PoS is maintaining explicit belief states that record the current world and unresolved requirements. It organizes the task-relevant world using entities, states, and relations: the cup, microwave, and cabinet are entities; whether the cup is hot and whether the microwave is closed are states; where the cup is located and who is holding it are expressed through relations. This representation keeps only the information that will affect actions, outcome judgments, and goal completion.

The second part is making "what is still missing" explicit. PoS divides unresolved requirements into two types: an Epistemic Gap refers to things that still need to be figured out, and an Achievement Gap refers to things that still need to be changed through action. When facing multiple unresolved requirements, PoS selects one Active Gap to prioritize at the moment; the agent can still see the complete task state, but its actions have a clear focus.

The third part is Belief Health and Trapping-Aware Recovery. The paper's setup starts detection after accumulating 8 verified transitions, and enters trapping diagnosis when belief health Hₜ ≤ 0.25. Recovery unfolds along two dimensions: behavioral patterns are classified as Static, Cycle, or Drift and judged in that order; for the blocked gap type, it compares the persistence of the epistemic gap and the achievement gap and selects the more stubborn one.

There is a design point here that is easy to overlook: the recovery constraints are handed to the original agent as additional conditions, and it continues to choose actions, with the current Active Gap remaining unchanged. After each new transition is verified, health is recalculated; if it is above the threshold, the temporary constraints are released, and if it is still trapped, the diagnosis and constraints are updated. Trapping detection, constraint generation, and post-recovery re-evaluation form a continuous closed loop.

Free for now

Read the full analysis

5 more sections of analysis, plus the full takeaway

Loading

Credibility boundary

All facts in this article come from a report by the WeChat public account "机器之心" (Machine Heart) on the paper "Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States." The method definition, component structure, threshold settings, case details, and experimental scale are clearly stated in the report and are treated as confirmed; the claim of achieving the highest overall metrics across 12 settings, along with the specific improvement margins, ablation values, and token overhead multiples, are all self-reported results from the paper, lacking third-party independent reproduction and first-party original source verification, and are treated as source_claim. The limitations self-described by the paper (dependence on the base model's ability to understand evidence, health not directly verifying the final answer, states possibly containing erroneous judgments caused by insufficient domain knowledge) and the unimplemented future directions likewise come from the paper's own account.

Sources for this version

  1. Agent一直在行动，任务为何没有进展？PoS识别并打破Belief Trapping机器之心

Primary report

机器之心

Primary source