{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.29098", "kind": "paper", "url": "https://arxiv.org/abs/2608.29098"}], "categories": ["General AI", "Multimodal", "Safety", "Safety & Trustworthiness"], "description": {"en": "Evaluates multimodal safety moderation via 5,000-instance held-out set with five-level ordered risk predictions and continuous scores across 15 harm categories."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.29098", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.29098", "title": "SafeAtlas-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2608.29098", "modality": null, "name": "SafeAtlas-Bench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"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_safeatlas-bench_6343d2e6", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "0346de8097c328b181a82e0f66be0c38502d91abb5efc377199b1835ba6d920f", "relation": "evaluates_only", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-01T06:28:51.098812Z", "source_url": "https://arxiv.org/abs/2608.29098"}, "publisher": null, "released": "2026-08-29", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.29098", "source_url": "https://arxiv.org/abs/2608.29098"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-29098", "source": "claire_radar", "source_benchmark_id": "2608.29098", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Machine Generated", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/zrwang1211/SafeAtlas-VL", "githubScope": "benchmark_repo", "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 9, "hfPaperUrl": "https://huggingface.co/papers/2608.29098", "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", "lastSuccessfulDate": "2026-09-29", "state": "stale"}}, "status": "stale"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "High-priority multimodal safety topic with a substantial dataset and guard model release should draw attention from both industry and academia.", "score": 70}, "capabilities": [], "confidence": "Medium", "construction": "Original Synthetic", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-01T06:28:51.098812Z", "inputHash": "ad95dccece041607ee817bdc6f06a6c924b2a14064f8922fdc16caea08f11083", "model": "deepseek-v4-pro"}, "curation": {"canonicalNameEvidence": "We also construct SafeAtlas-Bench, a held-out set of 5,000 instances for evaluating five-level predictions and continuous risk scores.", "canonicalNameSource": "abstract", "decisionReason": "Explicitly named benchmark with held-out set, five-level scoring, and released code, data, and models.", "model": "deepseek-v4-pro", "reviewedAt": "2026-09-01T06:28:51.098812Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates multimodal safety moderation via 5,000-instance held-out set with five-level ordered risk predictions and continuous scores across 15 harm categories.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "evaluation protocol evidence", "named entity is a model, method, framework, or dataset used in evaluation"], "snippet": "We also construct SafeAtlas-Bench, a held-out set of 5,000 instances for evaluating five-level predictions and continuous risk scores."}, "familyId": "bmf_123ea58abfe9", "firstSeenAt": "2026-09-01", "heat": null, "id": "bm_safeatlas-bench_6343d2e6", "indexedAt": "2026-09-01T06:28:29.895112Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2608.29098", "pdf": "https://arxiv.org/pdf/2608.29098", "project": null, "report": "https://arxiv.org/abs/2608.29098"}, "metrics": [], "motivation": "Multimodal safety moderation requires distinguishing risks arising from visual content, user intent, and assistant behavior.", "name": "SafeAtlas-Bench", "oneLine": "Evaluates multimodal safety moderation via 5,000-instance held-out set with five-level ordered risk predictions and continuous scores across 15 harm categories.", "paperTitle": "SafeAtlas-VL: Beyond Binary Multimodal Safety with Large-Scale Data and Guard Models", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.29098", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-01T06:28:29.895112Z"}, "publications": [], "ranking": {"90d": {"confidence": "Low", "coverage": 0.15, "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": 0.4068543738121793, "value": 9}}, "confidence": "Low", "coverage": 0.15, "rank": null, "score": 49}, "method": "attention-ranking-v13", "rank": null, "score": 49}}, "readiness": "Paper only", "recognitionConfidence": 0.55, "relation": "evaluates_only", "releasedAt": "2026-08-29", "source": {"authors": ["Zongrui Wang", "Xiangyang Zhu", "Sicheng Wang", "Han Wang", "Dingyi Rong", "Zeyu Zhang", "Chunyi Li", "Yue Shi", "Kaiwei Zhang", "Zicheng Zhang", "Yuan Tian", "Qi Jia", "Yan Teng", "Wei Sun", "Ning Liu", "Guangtao Zhai"], "categories": ["cs.AI", "cs.CV"], "id": "2608.29098", "title": "SafeAtlas-VL: Beyond Binary Multimodal Safety with Large-Scale Data and Guard Models", "type": "arxiv", "url": "https://arxiv.org/abs/2608.29098"}, "sourceUpdatedAt": "2026-08-29T07:08:01Z", "topics": ["Multimodal", "Safety"], "venueAttempts": [], "whyItMatters": "Enables nuanced, comparable safety assessment beyond binary judgments, improving risk detection in multimodal interactions."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
