Title: StepFinder: A Temporal Semantic Framework for Failure Attribution in Multi-Agent Systems

Abstract:

Large language model (LLM)-driven multi-agent systems have demonstrated impressive collaborative potential when tackling intricate, multi-stage tasks. Nevertheless, these systems are extremely vulnerable to errors occurring in individual steps, which can ripple through agent interactions and trigger cascading failures. To enhance system reliability and comprehend the origins of these breakdowns, the field has adopted failure attribution—a process designed to automatically pinpoint the specific root cause step responsible for a failure. Current approaches predominantly depend on LLMs to analyze original execution trajectories. However, this reliance leads to significant inference costs and latency, while also being susceptible to interference from redundant and noisy logs, often hindering the accurate identification of the true root cause.

To overcome these limitations, we introduce StepFinder, a lightweight framework for failure attribution. In our approach, LLMs are utilized exclusively during the feature construction stage to transform execution logs into temporal semantic sequences. Following this, a parameter-efficient integration of attention mechanisms and temporal modeling modules is employed to capture both the sequential progression and cross-step dependencies within the trajectories. The framework further refines step-level error scores using multi-scale differences and position bias to ensure precise root cause identification.

Evaluations on the Who&When benchmark reveal that StepFinder surpasses existing LLM-based methods in step-level failure attribution. Moreover, it delivers substantially greater inference efficiency, cutting inference time by 79% compared to the most efficient LLM-based alternative, with zero overhead associated with text generation. The source code for StepFinder is publicly accessible at https://github.com/taiyu-zhu/StepFinder.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC