{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2609.08155", "kind": "paper", "url": "https://arxiv.org/abs/2609.08155"}], "authors": ["Wentao Li", "Yibo Wu", "Yizhe Chen", "Ruixuan Chen", "Jiangjie Qiu", "Yijun Li", "Zhao Leyi", "Xiaonan Wang"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.08155", "source_url": "https://arxiv.org/abs/2609.08155"}, "categories": ["General AI", "Vision & 3D", "cs.CV"], "description": {"en": "Evaluates generation of editable Python programs that preserve presentation of scientific figure panels, using a 337-panel test set across 31 chart subtypes, five domains, and three complexity levels."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.08155", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.08155", "title": "SciFigure2Code"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Tsinghua University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2609.08155", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2609.08155", "source_url": "https://openalex.org/W7212036227"}, "key": "claire-radar:2609.08155", "modality": null, "name": "SciFigure2Code", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "High", "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_scifigure2code_b5a28dad", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "8f5405d55cae1fb66f208aa356e9b5dd4b45608b7c0e83a2b5b128608e91140c", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-09T06:27:57.312318Z", "source_url": "https://arxiv.org/abs/2609.08155"}, "publisher": null, "released": "2026-09-08", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.08155", "source_url": "https://arxiv.org/abs/2609.08155"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-08155", "source": "claire_radar", "source_benchmark_id": "2609.08155", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Vision & 3D", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2609.08155", "observedAt": "2026-10-04T03:02:42.754940Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "state": "unavailable"}}, "status": "unavailable"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The benchmark addresses a novel task in scientific figure understanding with a large curated dataset and results for 14 models, likely to draw interest from researchers in multimodal code generation and scientific document AI.", "score": 62}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-09T06:27:57.312318Z", "inputHash": "841c5f79b12bcf3bf10698cb1c923e656ad1286f05fb04c213d3ccdb6a5b46d6", "model": "deepseek-v4-pro"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "SciFigure2Code: An AI-Reconstructed Benchmark for Scientific Figure-to-Code", "canonicalNameSource": "paper_title", "decisionReason": "The benchmark defines a concrete evaluation object (generating Python code for figure presentation), a scoring setup, and includes a public test set with reported model scores, satisfying repeatable evaluation and public reuse path.", "evaluationEvidence": {"reuse": "We introduce SciFigure2Code, an AI-reconstructed benchmark", "scoring": "Across 14 zero-shot models in image-only and caption-assisted settings, execution, multi-component layouts, axes, legends, and scientific labels remain weak. Claude Opus 4.7 achieves the highest image-only Overall score, Claude Opus 4.6 leads caption-assisted reconstruction", "task": "generating editable Python programs that preserve how a scientific panel is arranged and read"}, "model": "deepseek-v4-pro", "reviewedAt": "2026-09-09T06:27:57.312318Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates generation of editable Python programs that preserve presentation of scientific figure panels, using a 337-panel test set across 31 chart subtypes, five domains, and three complexity levels.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We introduce SciFigure2Code, an AI-reconstructed benchmark that instead evaluates presentation recovery: generating editable Python programs that preserve how a scientific panel is arranged and read."}, "familyId": "bmf_ad7106a0904a", "firstSeenAt": "2026-09-09", "heat": null, "id": "bm_scifigure2code_b5a28dad", "indexedAt": "2026-09-09T06:18:29.794093Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2609.08155", "pdf": "https://arxiv.org/pdf/2609.08155", "project": null, "report": "https://arxiv.org/abs/2609.08155"}, "metrics": [], "motivation": "Scientific figures are the interface through which research claims are inspected and reused, but final published panels rarely expose the data or plotting code that produced them.", "name": "SciFigure2Code", "oneLine": "Evaluates generation of editable Python programs that preserve presentation of scientific figure panels, using a 337-panel test set across 31 chart subtypes, five domains, and three complexity levels.", "paperTitle": "SciFigure2Code: An AI-Reconstructed Benchmark for Scientific Figure-to-Code", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.08155", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-09T06:18:29.794093Z"}, "publications": [], "ranking": {"30d": {"confidence": "Low", "coverage": 0.0, "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": 30, "historyDate": "2026-09-05", "rank": null, "score": null, "windowDays": 30}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": null, "value": null}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v13", "rank": null, "score": null}, "90d": {"confidence": "Low", "coverage": 0.0, "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": null, "value": null}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v13", "rank": null, "score": null}}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-09-08", "source": {"authors": ["Wentao Li", "Yibo Wu", "Yizhe Chen", "Ruixuan Chen", "Jiangjie Qiu", "Yijun Li", "Zhao Leyi", "Xiaonan Wang"], "categories": ["cs.CV"], "id": "2609.08155", "title": "SciFigure2Code: An AI-Reconstructed Benchmark for Scientific Figure-to-Code", "type": "arxiv", "url": "https://arxiv.org/abs/2609.08155"}, "sourceUpdatedAt": "2026-09-08T02:39:02Z", "topics": ["cs.CV"], "venueAttempts": [], "whyItMatters": "Provides an auditable testbed for evaluating agents that construct editable, visually faithful scientific figure presentations from static panels."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
