Title: Pinpointing Mistakes in Deep-Research Agents: Span-Level Error Localization in Agent Trajectories

Abstract: Deep-research agents tackle complex tasks by navigating extensive sequences involving search queries, tool utilization, evidence evaluation, and final answer formulation. While current evaluation metrics primarily assess success based on the final output, they fail to identify specific segments of the agent’s trajectory that compromise the reliability of the result. This study focuses on localizing errors at the span level within deep-research agents. We compiled a dataset of 2,790 real-world trajectories derived from three benchmarks, three distinct backbone models, and two agent frameworks. By transforming raw logs into semantic spans, we employed LLM-assisted expert review to annotate harmful error segments. These annotations formed the basis of TELBench, a new benchmark comprising 1,000 instances designed to distinguish error spans from routine exploration, unsuccessful searches, provisional hypotheses, and benign noise. Additionally, we introduce DRIFT, a claim-centric auditing system that monitors agent assertions, verifies their alignment with trajectory evidence, and flags spans where unsupported or contradictory claims disrupt the path to the answer. Experimental results across various model families and auditing frameworks demonstrate that DRIFT enhances span-level error localization and first-error accuracy by as much as 30 percentage points. Our findings offer a process-oriented perspective on the reliability of deep-research agents.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC