{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.02060", "kind": "paper", "url": "https://arxiv.org/abs/2606.02060"}, {"id": "gh:nju-link/drift", "kind": "repo", "url": "https://github.com/NJU-LINK/DRIFT"}], "authors": ["Jiaming Wang", "Ziteng Feng", "Jiangtao Wu", "Ruihao Li", "Qianqian Xie", "Yuxiang Ren", "He Zhu", "Xueming Han", "Fanyu Meng", "Junlan Feng", "Jiaheng Liu"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.02060", "source_url": "https://arxiv.org/abs/2606.02060"}, "categories": ["Agents", "General AI", "Language & Knowledge"], "description": {"en": "Evaluates span-level error localization in deep-research agent trajectories. TELBench comprises 1,000 instances with annotations of harmful error spans among normal exploration, failed searches, tentative hypotheses, and harmless noise, scored by span-level localization and first-error accuracy."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.02060", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.02060", "title": "TELBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Nanjing University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.02060", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.02060", "source_url": "https://openalex.org/W7163181016"}, "key": "claire-radar:2606.02060", "modality": null, "name": "TELBench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "High", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "unknown", "origin_record_id": "bm_telbench_d23c11db", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "24c0b4f699d153d9da1c174d66604866dfc33f498223d754487115b779f68e01", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.02060"}, "publisher": null, "released": "2026-06-01", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.02060", "source_url": "https://arxiv.org/abs/2606.02060"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-02060", "source": "claire_radar", "source_benchmark_id": "2606.02060", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/NJU-LINK/DRIFT", "githubScope": "benchmark_repo", "githubStars": 24, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 58, "hfPaperUrl": "https://huggingface.co/papers/2606.02060", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-09-30", "state": "stale"}}, "status": "fresh"}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "ad0404261eeff9c40b323be9628c5f11c01515f0877b2df8dc1b3200e78aa8fb", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates span-level error localization in deep-research agent trajectories. TELBench comprises 1,000 instances with annotations of harmful error spans among normal exploration, failed searches, tentative hypotheses, and harmless noise, scored by span-level localization and first-error accuracy.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "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."}, "familyId": "bmf_8c4c87f14d99", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [{"evidence": [{"basis": "existing-code-link", "linkedUrl": "https://github.com/NJU-LINK/DRIFT", "url": "https://github.com/NJU-LINK/DRIFT"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/NJU-LINK/DRIFT", "url": "https://huggingface.co/papers/2606.02060"}], "scope": "project_repo", "url": "https://github.com/NJU-LINK/DRIFT", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_telbench_d23c11db", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/NJU-LINK/DRIFT", "data": null, "hfPaper": "https://huggingface.co/papers/2606.02060", "paper": "https://arxiv.org/abs/2606.02060", "pdf": "https://arxiv.org/pdf/2606.02060", "project": null, "report": "https://arxiv.org/abs/2606.02060"}, "metrics": [], "motivation": "Deep-research agents solve tasks through long trajectories of search, tool use, evidence inspection, and answer synthesis.", "name": "TELBench", "oneLine": "Evaluates span-level error localization in deep-research agent trajectories. TELBench comprises 1,000 instances with annotations of harmful error spans among normal exploration, failed searches, tentative hypotheses, and harmless noise, scored by span-level localization and first-error accuracy.", "paperTitle": "Where Do Deep-Research Agents Go Wrong? Span-Level Error Localization in Agent Trajectories", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.02060", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "NJU-LINK", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/NJU-LINK/DRIFT"}], "ranking": {}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-01", "source": {"authors": ["Jiaming Wang", "Ziteng Feng", "Jiangtao Wu", "Ruihao Li", "Qianqian Xie", "Yuxiang Ren", "He Zhu", "Xueming Han", "Fanyu Meng", "Junlan Feng", "Jiaheng Liu"], "categories": ["cs.AI"], "id": "2606.02060", "title": "Where Do Deep-Research Agents Go Wrong? Span-Level Error Localization in Agent Trajectories", "type": "arxiv", "url": "https://arxiv.org/abs/2606.02060"}, "sourceUpdatedAt": "2026-06-02T10:30:22Z", "topics": ["Agents"], "venueAttempts": [], "whyItMatters": "Final-answer evaluation does not reveal which trajectory steps make answers unreliable. This benchmark enables process-level reliability assessment and comparison of error localization methods for deep-research agents."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
