{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2605.30434", "kind": "paper", "url": "https://arxiv.org/abs/2605.30434"}, {"id": "gh:zjunlp/datamind", "kind": "repo", "url": "https://github.com/zjunlp/DataMind"}], "authors": ["Kewei Xu", "Xiaoben Lu", "Shuofei Qiao", "Zihan Ding", "Haoming Xu", "Lei Liang", "Ningyu Zhang"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2605.30434", "source_url": "https://arxiv.org/abs/2605.30434"}, "categories": ["General AI", "Language & Knowledge", "cs.LG"], "description": {"en": "LongDS-Bench evaluates long-horizon, multi-turn data analysis tasks where agents must maintain, update, restore, and compose evolving analytical states. It comprises 68 tasks from real-world Kaggle notebooks spanning 2,225 turns across six domains, with an average dependency span of 11.3 turns."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2605.30434", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2605.30434", "title": "LongDS-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2605.30434", "modality": null, "name": "LongDS-Bench", "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_longds-bench_2436ff1e", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "82df533de1500c9b64b75ce698d2d19fd0929c6d9b5f9e64a48db9600fcb9334", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2605.30434"}, "publisher": null, "released": "2026-05-28", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2605.30434", "source_url": "https://arxiv.org/abs/2605.30434"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2605-30434", "source": "claire_radar", "source_benchmark_id": "2605.30434", "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/zjunlp/DataMind", "githubScope": "benchmark_repo", "githubStars": 147, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 18, "hfPaperUrl": "https://huggingface.co/papers/2605.30434", "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": "aa20c40037c63cfe5d24156a4594e58fc804546ab10e82a764151d14dcbfc87c", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "LongDS-Bench evaluates long-horizon, multi-turn data analysis tasks where agents must maintain, update, restore, and compose evolving analytical states. It comprises 68 tasks from real-world Kaggle notebooks spanning 2,225 turns across six domains, with an average dependency span of 11.3 turns.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We introduce LongDS, a benchmark for long-horizon, multi-turn data analysis where agents must maintain, update, restore, and compose evolving analytical states."}, "familyId": "bmf_95132b949b16", "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/zjunlp/DataMind", "url": "https://github.com/zjunlp/DataMind"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/zjunlp/DataMind", "url": "https://huggingface.co/papers/2605.30434"}, {"basis": "linked-from-report", "linkedUrl": "https://github.com/zjunlp/DataMind.", "url": "https://arxiv.org/abs/2605.30434"}, {"basis": "linked-from-report", "linkedUrl": "https://github.com/zjunlp/DataMind", "url": "https://arxiv.org/abs/2605.30434"}], "scope": "project_repo", "url": "https://github.com/zjunlp/DataMind", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_longds-bench_2436ff1e", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/zjunlp/DataMind", "data": null, "hfPaper": "https://huggingface.co/papers/2605.30434", "paper": "https://arxiv.org/abs/2605.30434", "pdf": "https://arxiv.org/pdf/2605.30434", "project": null, "report": "https://arxiv.org/abs/2605.30434"}, "metrics": [], "motivation": "Real-world data analysis is inherently iterative, yet existing benchmarks mostly evaluate isolated or short interactive tasks, leaving agents' ability to track evolving analytical context over long horizons untested.", "name": "LongDS-Bench", "oneLine": "LongDS-Bench evaluates long-horizon, multi-turn data analysis tasks where agents must maintain, update, restore, and compose evolving analytical states. It comprises 68 tasks from real-world Kaggle notebooks spanning 2,225 turns across six domains, with an average dependency span of 11.3 turns.", "paperTitle": "LongDS-Bench: On the Failure of Long-Horizon Agentic Data Analysis", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2605.30434", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "Zhejiang University NLP Lab (ZJUNLP)", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/zjunlp/DataMind"}], "ranking": {}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-05-28", "source": {"authors": ["Kewei Xu", "Xiaoben Lu", "Shuofei Qiao", "Zihan Ding", "Haoming Xu", "Lei Liang", "Ningyu Zhang"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.MA"], "id": "2605.30434", "title": "LongDS-Bench: On the Failure of Long-Horizon Agentic Data Analysis", "type": "arxiv", "url": "https://arxiv.org/abs/2605.30434"}, "sourceUpdatedAt": "2026-05-28T18:00:20Z", "topics": ["cs.LG"], "venueAttempts": [], "whyItMatters": "Existing benchmarks focus on isolated or short interactive tasks, leaving long-horizon analytical state management untested. This benchmark reveals a critical bottleneck in agent performance, where errors concentrate in later turns and additional interaction steps do not reliably improve accuracy, aiding development of more reliable agentic systems."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
