跳转至

从序列到结构：面向大模型智能体的关系不确定性传播

文章背景与核心概要

在复杂的交互式环境中部署大语言模型（LLM）智能体时，可靠的不确定性量化（UQ）至关重要。传统的不确定性量化方法通常依赖于孤立的局部信号（如 Token 概率、预测熵或单步置信度），这些方法难以捕捉长距离依赖，无法有效评估在执行轨迹中不断累积的错误。因此，这些方法往往无法识别那些根源于最终答案产生前数个推理或交互步骤的失败。

为了解决这一局限性，本文提出了 RUPA（Relational Uncertainty Propagation for Agents），这是一个专为 LLM 智能体设计的轨迹级不确定性量化框架。RUPA 将执行历史建模为有向轨迹图，其中推理状态、工具交互和环境反馈作为节点，通过时间和语义依赖边连接。通过在图结构上传播不确定性，该框架能够捕捉执行风险如何在交互步骤之间累积和传递。实验结果表明，RUPA 在多个代表性基准测试中均优于现有方法，为长程智能体任务的可靠执行提供了坚实基础。

📌 摘要 (Summary)

可靠的不确定性量化（UQ）对于在复杂交互环境中部署大语言模型（LLM）智能体至关重要。传统 UQ 方法主要依赖于局部信号，如 Token 概率、预测熵或单步置信度，因此忽略了错误在执行轨迹中累积的长距离依赖关系。结果，它们可能无法识别那些在最终答案产生前数个推理或交互步骤中就已埋下失败隐患的情况。

为了解决这一局限性，作者引入了 RUPA (Relational Uncertainty Propagation for Agents)，这是一个专门为 LLM 智能体设计的轨迹级 UQ 框架。

RUPA 的核心亮点：

  • 轨迹图建模： RUPA 将执行历史表示为有向轨迹图，其中推理状态、工具交互和环境反馈作为节点，通过时间和语义依赖边进行连接。
  • 不确定性传播： 它在图结构上传播不确定性，以捕捉执行风险如何在交互步骤之间累积和传递。
  • 综合置信度评分： 将传播后的信号与轨迹级行为特征及目标对齐细节相结合，从而为整个智能体轨迹制定稳健的置信度估计。
  • 实证验证： 在代表性基准测试（\(\tau\)-2、Terminal-Bench-2 和 GAIA）上，使用 6 个跨不同模型系列的开源 LLM 进行了评估。结果表明，RUPA 通过提供更优的不确定性估计、实现更早的故障检测以及增强不确定性引导下的智能体性能，始终优于基准 UQ 方法。

Reliable uncertainty quantification (UQ) is critical for safely deploying Large Language Model (LLM) agents in complex, interactive environments. Traditional UQ methods generally rely on isolated local signals—such as token probabilities, predictive entropy, or per-step confidence—which fail to capture long-range dependencies where errors accumulate throughout an execution trajectory. Consequently, these methods often miss failures whose root causes occurred multiple reasoning or interaction steps prior to the final answer.

To address this limitation, the authors introduce RUPA (Relational Uncertainty Propagation for Agents), a trajectory-level UQ framework designed specifically for LLM agents.

Key Highlights of RUPA:

  • Trajectory Graph Modeling: RUPA represents execution histories as directed trajectory graphs where reasoning states, tool interactions, and environment feedback serve as nodes connected by temporal and semantic dependency edges.
  • Uncertainty Propagation: It propagates uncertainty across the graph structure to capture how execution risks accumulate and transfer between interaction steps.
  • Comprehensive Confidence Scoring: The propagated signals are merged with trajectory-level behavioral features and goal-alignment details to formulate a robust confidence estimate for the entire agent trajectory.
  • Empirical Validation: Evaluated across representative benchmarks (\(\tau\)-2, Terminal-Bench-2, and GAIA) using 6 open-source LLMs across various model families. Results demonstrate that RUPA consistently outperforms baseline UQ approaches by yielding superior uncertainty estimates, enabling earlier failure detection, and enhancing uncertainty-guided agent performance.

📄 论文摘要 (Abstract)

可靠的不确定性量化（UQ）对于在复杂交互环境中部署大语言模型（LLM）智能体至关重要。现有的 UQ 方法主要依赖于局部信号，如 Token 概率、预测熵或单步置信度，因此忽略了错误在执行轨迹中累积的长距离依赖关系。结果，它们可能无法识别那些在最终答案产生前数个推理或交互步骤中就已埋下失败隐患的情况。我们提出了 RUPA（Relational Uncertainty Propagation for Agents），这是一个面向 LLM 智能体的轨迹级 UQ 框架。RUPA 将执行历史表示为有向轨迹图，其中推理状态、工具交互和环境反馈作为节点，通过时间和语义依赖边进行连接。随后，它在图结构上传播不确定性，以捕捉执行风险如何在交互步骤之间累积和传递。传播后的信号与轨迹级行为特征及目标对齐信息相结合，从而为整个智能体轨迹生成置信度估计。我们在代表性智能体基准测试（包括 \(\tau\)-2、Terminal-Bench-2 和 GAIA）上评估了 RUPA，使用了跨多个模型系列的 6 个开源 LLM。实验结果表明，RUPA 通过提供更准确的不确定性估计、实现更早的故障检测以及改善不确定性引导下的智能体执行，始终优于现有的 UQ 方法。这些结果证明，显式建模关系依赖对于长程 LLM 智能体的可靠 UQ 至关重要，为可信智能体执行提供了实践基础。

Reliable uncertainty quantification (UQ) is essential for deploying large language model (LLM) agents in complex interactive environments. Existing UQ methods largely rely on local signals, such as token probabilities, predictive entropy, or per-step confidence, and therefore overlook the long-range dependencies through which errors accumulate across an execution trajectory. As a result, they may fail to identify agent failures whose causes originate several reasoning or interaction steps before the final answer. We propose RUPA (Relational Uncertainty Propagation for Agents), a trajectory-level UQ framework for LLM agents. RUPA represents an execution history as a directed trajectory graph in which reasoning states, tool interactions, and environment feedback are nodes connected by temporal and semantic dependency edges. It then propagates uncertainty over this graph to capture how execution risk accumulates and transfers across interaction steps. The propagated signal is combined with trajectory-level behavioral features and goal-alignment information to produce a confidence estimate for the full agent trajectory. We evaluate RUPA on representative agent benchmarks, including \(\tau\)-2, Terminal-Bench-2, and GAIA, using 6 open-source LLMs spanning multiple model families. Experimental results show that RUPA consistently outperforms existing UQ methods by providing more accurate uncertainty estimates, enabling earlier failure detection, and improving uncertainty-guided agent execution across diverse agent tasks. These results demonstrate that explicitly modeling relational dependency is crucial to reliable UQ for long-horizon LLM agents, providing a practical foundation for trustworthy agent execution.

🔗 全文与资源 (Full-Text & Resources)

📚 参考文献与外部引用 (References & External Citations)