Where Do Deep-Research Agents Go Wrong? Span-Level Error Localization in Agent Trajectories
Fuente:
arXiv
Guardado en:
| Autores principales: | , , , , , , , , , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2026
|
| Materias: | |
| Acceso en línea: | |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866911741252730880 |
|---|---|
| author | Wang, Jiaming Feng, Ziteng Wu, Jiangtao Li, Ruihao Xie, Qianqian Ren, Yuxiang Zhu, He Han, Xueming Meng, Fanyu Feng, Junlan Liu, Jiaheng |
| author_facet | Wang, Jiaming Feng, Ziteng Wu, Jiangtao Li, Ruihao Xie, Qianqian Ren, Yuxiang Zhu, He Han, Xueming Meng, Fanyu Feng, Junlan Liu, Jiaheng |
| contents | Deep-research agents solve tasks through long trajectories of search, tool use, evidence inspection, and answer synthesis. Evaluation based on final answers shows whether an agent succeeds, but not which parts of the trajectory make the answer unreliable. We study span-level error localization for deep-research agents. We collect 2,790 real trajectories from two agent frameworks, three backbone models, and three benchmarks, convert raw logs into semantic spans, and annotate harmful error spans through LLM-assisted expert review. From these annotations, we build TELBench, a 1,000-instance benchmark for identifying error spans among normal exploration, failed searches, tentative hypotheses, and harmless noise. We further propose DRIFT, a claim-centric auditing framework that tracks agent claims, checks their support in trajectory evidence, and marks spans where unsupported or conflicting claims affect the answer path. Experiments across model families and auditing frameworks show that DRIFT improves span-level error localization and first-error accuracy by up to 30 percentage points. Our work provides a process-level view of reliability in deep-research agents. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2606_02060 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Where Do Deep-Research Agents Go Wrong? Span-Level Error Localization in Agent Trajectories Wang, Jiaming Feng, Ziteng Wu, Jiangtao Li, Ruihao Xie, Qianqian Ren, Yuxiang Zhu, He Han, Xueming Meng, Fanyu Feng, Junlan Liu, Jiaheng Artificial Intelligence Deep-research agents solve tasks through long trajectories of search, tool use, evidence inspection, and answer synthesis. Evaluation based on final answers shows whether an agent succeeds, but not which parts of the trajectory make the answer unreliable. We study span-level error localization for deep-research agents. We collect 2,790 real trajectories from two agent frameworks, three backbone models, and three benchmarks, convert raw logs into semantic spans, and annotate harmful error spans through LLM-assisted expert review. From these annotations, we build TELBench, a 1,000-instance benchmark for identifying error spans among normal exploration, failed searches, tentative hypotheses, and harmless noise. We further propose DRIFT, a claim-centric auditing framework that tracks agent claims, checks their support in trajectory evidence, and marks spans where unsupported or conflicting claims affect the answer path. Experiments across model families and auditing frameworks show that DRIFT improves span-level error localization and first-error accuracy by up to 30 percentage points. Our work provides a process-level view of reliability in deep-research agents. |
| title | Where Do Deep-Research Agents Go Wrong? Span-Level Error Localization in Agent Trajectories |
| topic | Artificial Intelligence |
| url | https://arxiv.org/abs/2606.02060 |