{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2609.05141", "kind": "paper", "url": "https://arxiv.org/abs/2609.05141"}, {"id": "gh:internlm/scidocbench", "kind": "repo", "url": "https://github.com/InternLM/SciDocBench"}], "authors": ["Shenxi Wu", "Yuhong Liu", "Haosong Zhang", "Tongjin Zou", "Yanxun Zhang", "Gaochang Chen", "Dun Liang", "Jiaqi Wang", "Zhecan James Wang", "Yuhang Zang", "Dahua Lin"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.05141", "source_url": "https://arxiv.org/abs/2609.05141"}, "categories": ["General AI", "Language & Knowledge", "cs.AI"], "description": {"en": "SciDocBench evaluates scientific document understanding with 124 expert-authored questions across seven capability groups and 19 subtasks in five scientific domains. Each question is tested under four conditions (English vs Chinese and all-images-first vs interleaved), yielding 496 instances. Performance is reported on a 100-point scale. Data and pipeline are released via a public repository."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.05141", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.05141", "title": "SciDocBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Chinese University of Hong Kong", "Shanghai Artificial Intelligence Laboratory", "New York University", "Fudan University", "Shanghai Jiao Tong University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2609.05141", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2609.05141", "source_url": "https://openalex.org/W7211903576"}, "key": "claire-radar:2609.05141", "modality": null, "name": "SciDocBench", "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": "false", "origin_record_id": "bm_scidocbench_baf9612d", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "7f4ab6970591bf9bf037caaa92aaf02e9bce9523eed430a4b3230663d597190a", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-name-audit-deferred", "reviewed_at": "2026-09-08T00:56:29.478814Z", "source_url": "https://arxiv.org/abs/2609.05141"}, "publisher": null, "released": "2026-09-04", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.05141", "source_url": "https://arxiv.org/abs/2609.05141"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-05141", "source": "claire_radar", "source_benchmark_id": "2609.05141", "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/InternLM/SciDocBench", "githubScope": "benchmark_repo", "githubStars": 6, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 0, "hfPaperUrl": "https://huggingface.co/papers/2609.05141", "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": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "Scientific document understanding is a niche research area with moderate potential audience; the release includes code and data that could attract specialized interest.", "score": 55}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-08T08:59:06.803818Z", "inputHash": "cd95e4804b2fd75f6a8082d9aeea17a8152889810e31301990769e573aef1051", "model": "deepseek-v4-flash"}, "curation": {"canonicalNameEvidence": "SciDocBench: A Workflow-Centered Benchmark and Data Pipeline for Scientific Document Understanding", "canonicalNameSource": "paper_title", "decisionReason": "The code link is unverified with no excerpt, and the abstract does not explicitly mention data or evaluator release; public reuse cannot be confirmed.", "model": "deepseek-v4-pro", "reviewedAt": "2026-09-08T00:56:29.478814Z", "state": "ai-name-audit-deferred"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "SciDocBench evaluates scientific document understanding with 124 expert-authored questions across seven capability groups and 19 subtasks in five scientific domains. Each question is tested under four conditions (English vs Chinese and all-images-first vs interleaved), yielding 496 instances. Performance is reported on a 100-point scale. Data and pipeline are released via a public repository.", "displayEligible": false, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "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 SciDocBench, a workflow-centered benchmark for scientific document understanding."}, "familyId": "bmf_c13354cb7309", "firstSeenAt": "2026-09-07", "heat": null, "id": "bm_scidocbench_baf9612d", "indexedAt": "2026-09-07T06:22:02.605340Z", "industrySectors": [], "links": {"code": "https://github.com/InternLM/SciDocBench", "data": null, "paper": "https://arxiv.org/abs/2609.05141", "pdf": "https://arxiv.org/pdf/2609.05141", "project": null, "report": "https://arxiv.org/abs/2609.05141"}, "metrics": [], "motivation": "Scientific papers require models to reason jointly over text, equations, figures, tables, code, and datasets while preserving the provenance of supporting evidence.", "name": "SciDocBench", "oneLine": "SciDocBench evaluates scientific document understanding with 124 expert-authored questions across seven capability groups and 19 subtasks in five scientific domains. Each question is tested under four conditions (English vs Chinese and all-images-first vs interleaved), yielding 496 instances. Performance is reported on a 100-point scale. Data and pipeline are released via a public repository.", "paperTitle": "SciDocBench: A Workflow-Centered Benchmark and Data Pipeline for Scientific Document Understanding", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.05141", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-07T06:22:02.605340Z"}, "publications": [], "ranking": {"90d": {"confidence": "Medium", "coverage": 0.7, "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": 0.2597440249823917, "value": 6}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.0, "value": 0}}, "confidence": "Medium", "coverage": 0.7, "rank": 282, "score": 29}, "method": "attention-ranking-v13", "rank": 282, "score": 29}}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-09-04", "source": {"authors": ["Shenxi Wu", "Yuhong Liu", "Haosong Zhang", "Tongjin Zou", "Yanxun Zhang", "Gaochang Chen", "Dun Liang", "Jiaqi Wang", "Zhecan James Wang", "Yuhang Zang", "and Dahua Lin"], "categories": ["cs.AI"], "id": "2609.05141", "title": "SciDocBench: A Workflow-Centered Benchmark and Data Pipeline for Scientific Document Understanding", "type": "arxiv", "url": "https://arxiv.org/abs/2609.05141"}, "sourceUpdatedAt": "2026-09-04T13:39:48Z", "topics": ["cs.AI"], "venueAttempts": [], "whyItMatters": "Existing benchmarks address scientific document understanding in isolated tasks, not realistic research workflows combining text, equations, figures, and cross-references. SciDocBench provides a structured evaluation, shows that strong systems remain limited, and connects diagnostics to scalable training data for scientific assistants."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
