{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.21928", "kind": "paper", "url": "https://arxiv.org/abs/2608.21928v1"}], "authors": ["Zhesheng Zhang", "Jiahao Lu", "Wei Liu", "Cong Pan", "Jianhua Yang", "Yixiang Chen", "Hongyuan Yu", "Mengqi Zhang", "Kailin Lyu", "Zhumin Chen", "Keji He"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.21928", "source_url": "https://arxiv.org/abs/2608.21928"}, "categories": ["Robotics & Embodied AI", "Safety"], "description": {"en": "3,024 instruction-scene examples organized as same-scene Safe/Unsafe contrastive pairs across hazard categories, based on international safety standards. Evaluates VLM safety reasoning under latent contextual risk."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.21928v1", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.21928v1", "title": "GuardianBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Shandong University", "National University of Singapore", "Nanjing University of Aeronautics and Astronautics", "Chinese Academy of Sciences", "Institute of Automation", "Xiaomi (China)"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.21928", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.21928", "source_url": "https://openalex.org/W7204165071"}, "key": "claire-radar:2608.21928", "modality": null, "name": "GuardianBench", "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": "true", "origin_record_id": "bm_guardianbench_debc3a47", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "fafec0dab3b00058b18747900d860935d7fab0855e42eaaaf93738f9ac8fca2b", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-08-26T06:15:55.708965Z", "source_url": "https://arxiv.org/abs/2608.21928v1"}, "publisher": null, "released": "2026-08-22", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.21928", "source_url": "https://arxiv.org/abs/2608.21928v1"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-21928", "source": "claire_radar", "source_benchmark_id": "2608.21928", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["Robotics & Embodied AI"], "area": "Robotics & Embodied AI", "attention": {"asOf": "2026-08-26", "attemptedAt": "2026-08-25T17:51:33.814358Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2608.21928", "observedAt": "2026-08-24T11:31:30.653116Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-08-25T17:51:33.814358Z", "state": "unavailable"}}, "status": "unavailable"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "Direct relevance to embodied AI safety may attract attention from safety-focused researchers.", "score": 50}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-26T06:20:35.651311Z", "inputHash": "89fe92b5cf418e1bfc88a593b72c69ccc4a097ba9118347e7a464353ac43dd24", "model": "deepseek-v4-flash"}, "curation": {"canonicalNameEvidence": "GuardianBench: A Same-Scene Instruction-Contrastive Benchmark for Latent Contextual Risk in Embodied AI", "canonicalNameSource": "paper_title", "decisionReason": "The paper defines a benchmark with a clear task, scoring contract, and released suite, though direct download links are absent from supplied metadata.", "model": "deepseek-v4-pro", "reviewedAt": "2026-08-26T06:15:55.708965Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "3,024 instruction-scene examples organized as same-scene Safe/Unsafe contrastive pairs across hazard categories, based on international safety standards. Evaluates VLM safety reasoning under latent contextual risk.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We introduce GuardianBench, an instruction-contrastive benchmark grounded in international safety standards that isolates this latent contextual risk through 3,024 instruction-scene examples organized as same-scene Safe/Unsafe contrastive pairs across various hazard categories."}, "familyId": "bmf_e7335f67edf2", "firstSeenAt": "2026-08-25", "heat": null, "id": "bm_guardianbench_debc3a47", "indexedAt": "2026-08-25T17:37:25.889905Z", "industrySectors": ["Robotics"], "links": {"code": null, "data": null, "paper": "http://arxiv.org/abs/2608.21928v1", "pdf": "https://arxiv.org/pdf/2608.21928v1", "project": null, "report": "http://arxiv.org/abs/2608.21928v1"}, "metrics": [], "motivation": "In embodied AI, safety risk can be latent: a benign instruction and a safe scene become hazardous only when composed.", "name": "GuardianBench", "oneLine": "3,024 instruction-scene examples organized as same-scene Safe/Unsafe contrastive pairs across hazard categories, based on international safety standards. Evaluates VLM safety reasoning under latent contextual risk.", "paperTitle": "GuardianBench: A Same-Scene Instruction-Contrastive Benchmark for Latent Contextual Risk in Embodied AI", "primaryDomain": "Robotics & Embodied AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "http://arxiv.org/abs/2608.21928v1", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-25T17:37:25.889905Z"}, "publications": [], "ranking": {"30d": {"confidence": "Low", "coverage": 0.0, "growth": {"components": {"githubStars": {"percentile": null, "value": null}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 30}, "level": {"components": {"githubStars": {"percentile": null, "value": null}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v11", "rank": null, "score": null}, "90d": {"confidence": "Low", "coverage": 0.0, "growth": {"components": {"githubStars": {"percentile": null, "value": null}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"percentile": null, "value": null}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v11", "rank": null, "score": null}}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-08-22", "source": {"authors": ["Zhesheng Zhang", "Jiahao Lu", "Wei Liu", "Cong Pan", "Jianhua Yang", "Yixiang Chen", "Hongyuan Yu", "Mengqi Zhang", "Kailin Lyu", "Zhumin Chen", "Keji He"], "categories": ["cs.AI", "cs.CL", "cs.RO"], "id": "2608.21928", "title": "GuardianBench: A Same-Scene Instruction-Contrastive Benchmark for Latent Contextual Risk in Embodied AI", "type": "arxiv", "url": "http://arxiv.org/abs/2608.21928v1"}, "sourceUpdatedAt": "2026-08-22T11:40:27Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "Exposes a critical failure mode in embodied AI safety—instruction-insensitive verdicts—and provides a controlled suite for improvement research."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
