arXiv cs.AITuesday · June 2, 2026FREE

FALAT: Tracing Failures in LLM Agent Trajectories via Dependency-Guided Search

A new diagnostic framework, FALAT (Failure Attribution in LLM Agent Trajectories), has been proposed by researchers to address the complex problem of identifying root causes of failures in large language model (LLM)-based agent systems. As LLM agents tackle increasingly intricate tasks involving multiple reasoning steps, tool calls, and inter-agent communication, pinpointing the exact step or agent responsible for a failure becomes challenging due to the propagation of errors. Mistakes can corrupt subsequent states, making later actions appear incorrect even if they are merely consequences of an earlier, decisive error. FALAT frames failure attribution as a dependency-guided search problem. The framework begins by constructing an "expectation" of how a given task should ideally be solved. This expectation is then used to pinpoint suspicious regions within an agent's actual trajectory where deviations from the ideal path might indicate an issue. Crucially, FALAT traces dependencies among various elements of the trajectory, including agent decisions, tool outputs, and inter-agent messages. This dependency analysis allows the system to differentiate between steps that genuinely introduce errors and those that simply inherit or propagate mistakes originating from an earlier point. The ultimate goal is to evaluate whether correcting a candidate problematic step would be sufficient to recover the expected successful outcome, thereby providing a precise attribution of the failure. This research, published on arXiv on June 2, 2026, aims to enhance the debuggability and reliability of sophisticated LLM agent applications.