{"record": {"aliases": ["Benchmarking Hybrid Deep Research Across Database Querying and Web Search"], "artifacts": [{"id": "arxiv:2609.09410", "kind": "paper", "url": "https://arxiv.org/abs/2609.09410"}, {"id": "gh:snowflake-ai-research/hybriddeepresearch", "kind": "repo", "url": "https://github.com/Snowflake-AI-Research/HybridDeepResearch"}, {"id": "hf:snowflake/hybriddeepresearch", "kind": "dataset", "url": "https://huggingface.co/datasets/Snowflake/HybridDeepResearch"}], "authors": ["Ruofan Wu", "Peiran Xu", "Xiaolong Li", "Fan Shu", "Soyoung Yoon", "Yite Wang", "Xiaodong Yu", "Boyi Liu", "Feng Yan", "Debiao Li", "Yuxiong He", "Zhewei Yao"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.09410", "source_url": "https://arxiv.org/abs/2609.09410"}, "categories": ["Factuality", "General AI", "Language & Knowledge", "cs.CL"], "description": {"en": "Evaluates deep-research agents on 380 tasks requiring both SQL database querying and web search to produce a complete end-to-end answer across SQL-to-Search, Search-to-SQL, and Parallel reasoning patterns."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.09410", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.09410", "title": "HybridDeepResearch"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["University of Houston", "University of California, Los Angeles", "University of Hong Kong", "Seoul National University", "Snowflake Inc. (United States)"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2609.09410", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2609.09410", "source_url": "https://openalex.org/W7212185320"}, "key": "claire-radar:2609.09410", "modality": null, "name": "HybridDeepResearch", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_benchmarking-hybrid-deep-research-across-d_5787d02b", "origin_source": "arxiv", "recognition_confidence": "0.7", "record_sha256": "d2a296adb067ee90731afaf0ae771e8b85ea1975914f970a1ccee5b8d8a48dd9", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-10T06:21:31.238194Z", "source_url": "https://arxiv.org/abs/2609.09410"}, "publisher": null, "released": "2026-09-08", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.09410", "source_url": "https://arxiv.org/abs/2609.09410"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-09410", "source": "claire_radar", "source_benchmark_id": "2609.09410", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Benchmarking Hybrid Deep Research Across Database Querying and Web Search"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-09-10", "attemptedAt": "2026-09-10T01:41:07.072503Z", "githubRepo": "https://github.com/Snowflake-AI-Research/HybridDeepResearch", "githubScope": "benchmark_repo", "githubStars": 1, "hfDailySubmittedAt": null, "hfDataset": "Snowflake/HybridDeepResearch", "hfDatasetDownloads": 22, "hfDatasetLikes": 0, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2609.09410", "observedAt": "2026-09-10T01:41:07.072503Z", "signalStatus": {"githubStars": {"observedAt": "2026-09-10T01:41:07.072503Z", "state": "fresh"}, "hfDatasetDownloads": {"observedAt": "2026-09-10T01:41:07.072503Z", "state": "fresh"}, "hfDatasetLikes": {"observedAt": "2026-09-10T01:41:07.072503Z", "state": "fresh"}, "hfPaperUpvotes": {"checkedAt": "2026-09-10T01:41:07.072503Z", "state": "unavailable"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The niche focus on hybrid SQL-web agent reasoning, explicit benchmark artifacts, and inclusion of prominent model results are likely to draw moderate initial attention.", "score": 75}, "capabilities": ["Factuality"], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-10T06:21:31.238194Z", "inputHash": "a545e6684e562b5b4631f483ee56cefc9d575d77a962b807908376683887b7ae", "model": "deepseek-v4-pro"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We introduce HybridDeepResearch, to our knowledge the first deep-research benchmark that requires both web search and SQL to form a complete, verifiable answer.", "canonicalNameSource": "abstract", "decisionReason": "The benchmark has a formally named release, clear task structure, explicit scoring rules, and publicly available code and data enabling external submissions and local evaluation.", "evaluationEvidence": {"reuse": "The submission expected submission artifact is a JSONL file keyed by instance_id, together with system metadata and a short method description. Before preparing an official submission, read docs/submission.md. The public release is intended for local development and iteration with read-only SQLite databases. Official held-out evaluation uses a governed Snowflake database.", "scoring": "Evaluation Scoring is end-to-end and binary. Partial progress on only the SQL or search leg does not count as a correct solution. - SQL2S and Parallel: concise text answers are compared against gold answers using a deterministic (temperature=0) BrowseComp-style LLM judge. - S2SQL: the submitted canonical SQL and gold SQL are executed read-only; their result tables are compared after column alignment.", "task": "The benchmark contains 380 tool-dependent tasks grounded in LiveSQLBench-Base-Lite databases and public web corpora, validated through automated checks and human review, and covering three reasoning patterns: SQL2S, S2SQL, and Parallel."}, "model": "deepseek-v4-pro", "reviewedAt": "2026-09-10T06:21:31.238194Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates deep-research agents on 380 tasks requiring both SQL database querying and web search to produce a complete end-to-end answer across SQL-to-Search, Search-to-SQL, and Parallel reasoning patterns.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We introduce HybridDeepResearch, to our knowledge the first deep-research benchmark that requires both web search and SQL to form a complete, verifiable answer."}, "familyId": "bmf_95ce4e777bff", "firstSeenAt": "2026-09-10", "heat": null, "id": "bm_benchmarking-hybrid-deep-research-across-d_5787d02b", "indexedAt": "2026-09-10T00:47:26.889471Z", "industrySectors": [], "links": {"code": "https://github.com/Snowflake-AI-Research/HybridDeepResearch", "data": "https://huggingface.co/datasets/Snowflake/HybridDeepResearch", "paper": "https://arxiv.org/abs/2609.09410", "pdf": "https://arxiv.org/pdf/2609.09410", "project": null, "report": "https://arxiv.org/abs/2609.09410"}, "metrics": [], "motivation": "While autonomous agents have made significant strides in \"deep research\" by iteratively navigating the open web to synthesize information, real-world problem-solving is rarely confined to a single environment.", "name": "HybridDeepResearch", "oneLine": "Evaluates deep-research agents on 380 tasks requiring both SQL database querying and web search to produce a complete end-to-end answer across SQL-to-Search, Search-to-SQL, and Parallel reasoning patterns.", "paperTitle": "Benchmarking Hybrid Deep Research Across Database Querying and Web Search", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.09410", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-10T00:47:26.889471Z"}, "publications": [], "publishers": [{"name": "Snowflake AI Research", "organizationType": "company-research-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/Snowflake-AI-Research/HybridDeepResearch"}], "ranking": {"30d": {"confidence": "Medium", "coverage": 0.7, "datasetDownloadRank": 39, "datasetRankPopulation": 41, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 30}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.10323645750017778, "value": 1}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": 0.32136554343715523, "value": 22}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Medium", "coverage": 0.7, "rank": 132, "score": 25}, "method": "attention-ranking-v13", "rank": 132, "score": 25}, "90d": {"confidence": "High", "coverage": 0.85, "datasetDownloadRank": 72, "datasetRankPopulation": 75, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.09355303223343062, "value": 1}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": 0.2599399034719878, "value": 22}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "High", "coverage": 0.85, "rank": 442, "score": 20}, "method": "attention-ranking-v13", "rank": 442, "score": 20}}, "readiness": "Runnable", "recognitionConfidence": 0.7, "relation": "introduces", "releasedAt": "2026-09-08", "source": {"authors": ["Ruofan Wu", "Peiran Xu", "Xiaolong Li", "Fan Shu", "Soyoung Yoon", "Yite Wang", "Xiaodong Yu", "Boyi Liu", "Feng Yan", "Debiao Li", "Yuxiong He", "Zhewei Yao"], "categories": ["cs.CL"], "id": "2609.09410", "title": "Benchmarking Hybrid Deep Research Across Database Querying and Web Search", "type": "arxiv", "url": "https://arxiv.org/abs/2609.09410"}, "sourceUpdatedAt": "2026-09-08T20:05:09Z", "topics": ["cs.CL"], "venueAttempts": [], "whyItMatters": "Fills the gap in benchmark coverage for cross-modal reasoning, measuring an agent's ability to preserve constraints when transferring evidence between structured databases and unstructured web text."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
