{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.12671", "kind": "paper", "url": "https://arxiv.org/abs/2606.12671"}, {"id": "hf:salartvqa/salart-vqa", "kind": "dataset", "url": "https://huggingface.co/datasets/salartvqa/SalArt-VQA"}], "authors": ["Ruotian Zhang", "Xiaoxiao Sun", "Junzhe Huang", "James Burgess", "Serena Yeung-Levy"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.12671", "source_url": "https://arxiv.org/abs/2606.12671"}, "categories": ["General AI", "Vision & 3D", "cs.CV"], "description": {"en": "SalArt-VQA is a closed-set visual question answering benchmark for fine-grained salient artifact understanding in AI-generated images. It includes 950 images and 3,681 human-authored multiple-choice questions covering artifact images, matched real references, and paired generated references. Four aligned question types evaluate presence detection, semantic localization, spatial grounding, and evidence-grounded defect identification, with reference splits for calibration and abstention testing."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.12671", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.12671", "title": "SalArt-VQA"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2606.12671", "modality": null, "name": "SalArt-VQA", "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_salart-vqa_f96d6eb3", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "caac31dc78be23980c62b18faef5e8ea63df621ce3852c584c004fa886c095c9", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.12671"}, "publisher": null, "released": "2026-06-10", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.12671", "source_url": "https://arxiv.org/abs/2606.12671"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-12671", "source": "claire_radar", "source_benchmark_id": "2606.12671", "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": "salartvqa/SalArt-VQA", "hfDatasetDownloads": 731, "hfDatasetLikes": 2, "hfPaperUpvotes": 1, "hfPaperUrl": "https://huggingface.co/papers/2606.12671", "observedAt": "2026-10-04T03:02:42.754940Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-01", "state": "stale"}, "hfDatasetLikes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-01", "state": "stale"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-01", "state": "stale"}}, "status": "stale"}, "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": "86c84fb08aa08659e330506ba3665d09266cff50f02d60d2a9f0bf8f8c6e6408", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "SalArt-VQA is a closed-set visual question answering benchmark for fine-grained salient artifact understanding in AI-generated images. It includes 950 images and 3,681 human-authored multiple-choice questions covering artifact images, matched real references, and paired generated references. Four aligned question types evaluate presence detection, semantic localization, spatial grounding, and evidence-grounded defect identification, with reference splits for calibration and abstention testing.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence", "public artifact URL"], "snippet": "To evaluate these behaviors directly, we introduce SalArt-VQA, a diagnostic benchmark for fine-grained SALient ARTifact understanding in AI-generated images."}, "familyId": "bmf_22505c6b720b", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_salart-vqa_f96d6eb3", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": "https://huggingface.co/datasets/salartvqa/SalArt-VQA", "hfPaper": "https://huggingface.co/papers/2606.12671", "paper": "https://arxiv.org/abs/2606.12671", "pdf": "https://arxiv.org/pdf/2606.12671", "project": null, "report": "https://arxiv.org/abs/2606.12671"}, "metrics": [], "motivation": "Vision-language models (VLMs) are increasingly used to detect whether AI-generated images contain visible artifacts, yet their ability to analyze such artifacts remains poorly understood.", "name": "SalArt-VQA", "oneLine": "SalArt-VQA is a closed-set visual question answering benchmark for fine-grained salient artifact understanding in AI-generated images. It includes 950 images and 3,681 human-authored multiple-choice questions covering artifact images, matched real references, and paired generated references. Four aligned question types evaluate presence detection, semantic localization, spatial grounding, and evidence-grounded defect identification, with reference splits for calibration and abstention testing.", "paperTitle": "SalArt-VQA: Diagnosing Whether VLMs Understand Salient Artifacts in Generated Images", "primaryDomain": "General AI", "publication": {"evidence": "Accepted to NeurIPS 2026, E&D Track (Oral). 23 pages, 7 figures, 7 tables. Dataset: https://huggingface.co/datasets/salartvqa/SalArt-VQA", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2606.12671", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "NeurIPS 2026, E&D Track (Oral)", "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "SalArt-VQA Benchmark Team", "organizationType": "community", "role": "benchmark-publisher", "sourceUrl": "https://huggingface.co/datasets/salartvqa/SalArt-VQA"}], "ranking": {}, "readiness": "Inspectable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-10", "source": {"authors": ["Xiaoxiao Sun", "Ruotian Zhang", "Junzhe Huang", "James Burgess", "Serena Yeung-Levy"], "categories": ["cs.CV"], "id": "2606.12671", "title": "SalArt-VQA: Diagnosing Whether VLMs Understand Salient Artifacts in Generated Images", "type": "arxiv", "url": "https://arxiv.org/abs/2606.12671"}, "sourceUpdatedAt": "2026-06-10T20:55:50Z", "topics": ["cs.CV"], "venueAttempts": [{"decisionRaw": "Accepted to NeurIPS 2026, E&D Track (Oral). 23 pages, 7 figures, 7 tables. Dataset: https://huggingface.co/datasets/salartvqa/SalArt-VQA", "evidence": [{"level": "author-claim", "observedAt": "2026-10-14T01:46:58.058199Z", "rawValue": "Accepted to NeurIPS 2026, E&D Track (Oral). 23 pages, 7 figures, 7 tables. Dataset: https://huggingface.co/datasets/salartvqa/SalArt-VQA", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2606.12671"}], "reviewStatus": "accepted", "venueName": "NeurIPS 2026, E&D Track (Oral)"}], "whyItMatters": "Image-level artifact detection accuracy can conceal failures in grounding and evidence use. This benchmark provides a fine-grained evaluation protocol that isolates specific failure modes, enabling comparison of VLMs on their ability to support artifact claims with local visual evidence. It offers practical value for developers selecting models for trustworthy artifact analysis in generated image workflows."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
