跳转至

衡量语言模型智能体中的跨任务行为一致性

文章背景与核心概要

当前的语言模型（LM）智能体评估方法过度依赖基于结果的指标（例如成功率），这些指标只能衡量智能体是否成功，却无法捕捉其行为的一致性。为了填补这一空白，本文引入了行为一致性指标（BCM），用于量化跨任务的行为一致性。通过分析六个LM智能体在软件 engineering 任务中的约9,000条执行轨迹，作者发现跨任务一致性与任务内一致性往往会发生背离：某些智能体系统在单个任务的重复尝试中表现出局部可复现性，但在不同任务之间却缺乏稳定策略（呈全局碎片化状态）。此外，一致性与成功率正交，成功率相近的系统可能在行为一致性上表现出巨大差异。

Current evaluations of language model (LM) agents rely heavily on outcome-based metrics (such as success rates), which measure whether an agent succeeds but fail to capture how consistently it behaves.

To address this gap, this paper introduces the Behavioral Consistency Metric (BCM) to quantify cross-task behavioral consistency. By analyzing roughly 9,000 execution trajectories from six LM agents on software engineering tasks, the authors reveal that: 1. Cross-task and within-task consistency diverge: Some agent systems are locally reproducible (behaving similarly during repeated attempts at a single task) yet globally fragmented (lacking a stable strategy across different tasks). 2. Consistency is orthogonal to success: Systems with comparable success rates can exhibit stark differences in behavioral consistency. 3. The capability gap persists: The consistency gap between frontier models and open-source models remains robust even when controlling for task difficulty.

Ultimately, BCM serves as a process-level reliability signal that complements traditional outcome metrics.

摘要 (Summary)

Agent evaluation relies almost entirely on outcome metrics such as success rate, which capture whether an agent succeeds but not how consistently it behaves. We argue that behavioral consistency across tasks is a distinct and measurable property, and we introduce the Behavioral Consistency Metric (BCM) to quantify it. BCM trains a model to predict task success from behavioral features of agent execution traces, derives a per-trajectory feature-attribution vector, and measures the mean pairwise similarity of these vectors within an agent system. Across roughly 9,000 trajectories from six language model agents on software engineering tasks, our central finding is that cross-task and within-task consistency are distinct axes that can diverge: some systems are locally reproducible, behaving similarly on repeated attempts at one task, yet globally fragmented, with no stable strategy across different tasks, while others are consistent at both scales. Prior work measures only same-task reproducibility and so cannot observe this separation. We further find that consistency is not reducible to success rate, since systems with comparable success can differ sharply in consistency, and that the frontier-versus-open-source consistency gap persists under a within-task control that holds task difficulty constant. We position BCM as a process-level reliability signal that complements outcome metrics, and we are explicit about the conditions under which it is meaningful.

核心发现与贡献 (Key Findings & Contributions)

  • 行为一致性指标 (BCM)： 一种原则性的方法，通过在智能体执行轨迹上训练预测模型、推导特征归因向量，并计算平均成对相似度来量化行为稳定性。
  • 一致性尺度的背离： 识别出局部可复现性（同任务）与全局一致性（跨任务）是两个解耦的维度。
  • 独立于成功率： 证明了达到相同总体成功率的智能体可能拥有截然不同的操作策略和可靠性特征。
  • 稳健的前沿与开源模型差距： 证实了前沿模型在性能和一致性上的优势在严格控制任务难度的情况下依然存在。
  • Behavioral Consistency Metric (BCM): A principled approach that trains predictive models on agent execution traces, derives feature-attribution vectors, and computes mean pairwise similarities to quantify behavioral stability.
  • The Divergence of Consistency Scales: Identifies that local reproducibility (same-task) and global consistency (cross-task) are decoupled dimensions.
  • Independence from Success Rates: Demonstrates that agents achieving the same overall success rate can possess drastically different operational strategies and reliability profiles.
  • Robust Frontier-Open Source Gap: Confirms that performance and consistency advantages of frontier models persist even under rigorous difficulty controls.