跳转至

文章背景与核心概要

在长程多步骤的工具使用任务中，大语言模型智能体经常面临一个关键挑战：即如何准确判断何时应该结束任务并提交最终答案，还是应该继续执行后续步骤。本文深入研究了“后期压力状态”（Late-Stage Pressure States）这一现象——在此状态下，智能体会倾向于提交那些表面上看起来精修且完整、但实际上关键约束仍未解决的过早最终答案。

为了解决这一问题，作者证明了可以通过线性探测器（Linear Probe）从智能体的隐藏状态中识别出这种后期压力状态。通过沿压力方向施加激活干预（Activation Interventions），研究表明直接改变这些隐藏状态可以有效影响智能体是选择继续调用工具还是提前提交。此外，通过控制上下文的操纵实验，作者发现增强约束清晰度和动作映射能够有效缓解这种压力。

基于这些深刻洞察，作者提出了“探测感知压力缓解”（Probe-Sensed Pressure Relief, PSPR）这一轻量级插件。该插件在处于中度压力时施加压力缓解方向的干预，而在面临高度压力风险时转向结构化组织。在多个长程基准测试上的实验表明，PSPR 能够持续显著地增强现有智能体框架的性能。

Polished but Unresolved: Identifying Late-Stage Pressure States in Long-Horizon Tool-Use Agents

Summary

Long-horizon tool-use agents frequently face a critical challenge: knowing when to finalize a task versus when to continue executing. This paper investigates late-stage pressure states, a phenomenon where agents are biased toward submitting premature final answers that appear polished and complete, despite crucial constraints remaining unresolved.

To address this, the authors demonstrate that late-stage pressure states can be identified from the agent's hidden states using a linear probe. By applying activation interventions along the pressure direction, they show that shifting these hidden states directly influences whether the agent continues tool use or submits early. Furthermore, controlled context manipulations reveal that constraint clarity and action mapping effectively mitigate this pressure.

Based on these insights, the authors propose Probe-Sensed Pressure Relief (PSPR), a lightweight plugin that applies pressure relief directions under moderate pressure and shifts to structured organization during high-pressure risks. Experiments across multiple long-horizon benchmarks demonstrate that PSPR consistently enhances existing agent frameworks.

长程工具使用智能体不仅需要搜索和规划，还需要决定何时进行收尾。我们研究了后期压力状态（late-stage pressure states），在这种状态下，智能体倾向于提交一个看起来完整且精修的最终答案，而关键约束条件实际上并未得到解决。我们首先训练了一个线性探测器，证明可以从智能体的隐藏状态中识别出这种压力状态。随后，我们沿着这个压力方向应用激活干预，发现改变隐藏状态既会改变压力得分，也会影响智能体是继续使用工具还是提前提交。通过控制上下文操纵，我们进一步发现压力可以通过约束清晰度和动作映射得到缓解。基于这些发现，我们提出了探针感知压力缓解（Probe-Sensed Pressure Relief, PSPR）插件，它在适度压力下应用轻量级压力缓解方向，在高压力风险下转向结构化组织。在多个长程基准测试上的实验表明，我们的方法持续增强了现有的智能体方法。

Metadata

  • arXiv ID: arXiv:2609.00823 [cs.AI]
  • Subject Categories: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
  • Conference Acceptance: Accepted at the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026)
  • Submission Date: September 1, 2026
  • Authors: Haoyang Chen, Yi Liu, Jianzhi Shao, Xiaozhou Xu, Zhe Sun, Wei Hu
  • arXiv ID: arXiv:2609.00823 [cs.AI]
  • 学科分类: 人工智能 (cs.AI)；计算与语言 (cs.CL)
  • 会议录用: 已被 2026年自然语言处理经验方法国际会议 (EMNLP 2026) 录用
  • 提交日期: 2026年9月1日
  • 作者: Haoyang Chen, Yi Liu, Jianzhi Shao, Xiaozhou Xu, Zhe Sun, Wei Hu

Abstract

Long-horizon tool-use agents need not only to search and plan, but also to decide when to finalize. We study late-stage pressure states, in which an agent is biased toward submitting a final answer that appears complete and polished while key constraints remain unresolved. We first train a linear probe to show that this pressure state is identifiable from the agent's hidden states. Then, we use activation interventions along this pressure direction and find that shifting the hidden states changes both the pressure score and whether the agent continues tool use or submits early. Through controlled context manipulations, we further see that the pressure is mitigated by constraint clarity and action mapping. Based on these findings, we propose Probe-Sensed Pressure Relief (PSPR), a plugin that applies lightweight pressure relief direction under moderate pressure and moves to structured organization under high pressure risk. Experiments on multiple long-horizon benchmarks show that our method consistently strengthens existing agent methods.

长程工具使用智能体不仅需要搜索和规划，还需要决定何时进行收尾。我们研究了后期压力状态（late-stage pressure states），在这种状态下，智能体倾向于提交一个看起来完整且精修的最终答案，而关键约束条件实际上并未得到解决。我们首先训练了一个线性探测器，证明可以从智能体的隐藏状态中识别出这种压力状态。随后，我们沿着这个压力方向应用激活干预，发现改变隐藏状态既会改变压力得分，也会影响智能体是继续使用工具还是提前提交。通过控制上下文操纵，我们进一步发现压力可以通过约束清晰度和动作映射得到缓解。基于这些发现，我们提出了探针感知压力缓解（Probe-Sensed Pressure Relief, PSPR）插件，它在适度压力下应用轻量级压力缓解方向，在高压力风险下转向结构化组织。在多个长程基准测试上的实验表明，我们的方法持续增强了现有的智能体方法。