{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.22873", "kind": "paper", "url": "https://arxiv.org/abs/2606.22873"}, {"id": "gh:inclusionai/sing-guard", "kind": "repo", "url": "https://github.com/inclusionAI/Sing-Guard"}], "authors": ["SingGuard Team"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.22873", "source_url": "https://arxiv.org/abs/2606.22873"}, "categories": ["Biology & Drug Discovery", "Finance", "Multimodal", "Reasoning"], "description": {"en": "A multimodal guardrail evaluation covering multimodal question answering, adversarial attacks, dynamic policy rules, and cross-modal joint risks."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.22873", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.22873", "title": "SingGuard-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2606.22873", "modality": null, "name": "SingGuard-Bench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-09-25T04:40:22Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_singguard-bench_67c1078d", "origin_source": "arxiv", "recognition_confidence": "0.6", "record_sha256": "4fdc9623f3332455c11a2b6f9f67c47140332362ee284f84502a93183fb4f890", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T18:36:35.723890Z", "source_url": "https://arxiv.org/abs/2606.22873"}, "publisher": null, "released": "2026-06-22", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.22873", "source_url": "https://arxiv.org/abs/2606.22873"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-22873", "source": "claire_radar", "source_benchmark_id": "2606.22873", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["Biology & Drug Discovery", "Finance"], "area": "Multimodal", "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "The combination of multimodal safety, dynamic-policy evaluation, and a released codebase may attract near-term attention.", "score": 72}, "capabilities": ["Reasoning"], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T18:36:35.723890Z", "inputHash": "1dd449929242bab3ca304f70acdf81619bcab8d752ba66370ebb13c827612805", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We also introduce \\textbf{SingGuard-Bench}", "canonicalNameSource": "abstract", "decisionReason": "Publish: the abstract formally names the benchmark, identifies its evaluation settings and metrics, and supplies a public code path.", "evaluationEvidence": {"reuse": "Our code is available at https://github.com/inclusionAI/Sing-Guard.", "scoring": "SingGuard achieves state-of-the-art average F1 in every family. Dynamic-rule evaluation further shows improved policy-following accuracy from 0.6465 to 0.7415", "task": "a multimodal guardrail benchmark with 56{,}340 examples spanning 80+ fine-grained risk types across multimodal QA, adversarial attack, and dynamic-rule evaluation settings"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T18:36:35.723890Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A multimodal guardrail evaluation covering multimodal question answering, adversarial attacks, dynamic policy rules, and cross-modal joint risks.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "public artifact URL"], "snippet": "We also introduce \\textbf{SingGuard-Bench}, a multimodal guardrail benchmark with 56{,}340 examples spanning 80+ fine-grained risk types across multimodal QA, adversarial attack, and dynamic-rule evaluation settings, including cross-modal joint-risk cases where each modality is harmless in isolation but their composition implies unsafe intent."}, "familyId": "bmf_d84092fcd4cb", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_singguard-bench_67c1078d", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Pharma & Biotech", "Financial Services"], "links": {"code": "https://github.com/inclusionAI/Sing-Guard", "data": null, "paper": "https://arxiv.org/abs/2606.22873", "pdf": "https://arxiv.org/pdf/2606.22873", "project": null, "report": "https://arxiv.org/abs/2606.22873"}, "metrics": [], "motivation": "Vision-language models (VLMs) are increasingly deployed in consumer, medical, financial, and enterprise applications.", "name": "SingGuard-Bench", "oneLine": "A multimodal guardrail evaluation covering multimodal question answering, adversarial attacks, dynamic policy rules, and cross-modal joint risks.", "paperTitle": "SingGuard: A Policy-Adaptive Multimodal LLM Guardrail with Dynamic Reasoning", "primaryDomain": "Biology & Drug Discovery", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.22873", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Runnable", "recognitionConfidence": 0.6, "relation": "introduces", "releasedAt": "2026-06-22", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["SingGuard Team"], "categories": ["cs.CV", "cs.CL"], "id": "2606.22873", "title": "SingGuard: A Policy-Adaptive Multimodal LLM Guardrail with Dynamic Reasoning", "type": "arxiv", "url": "https://arxiv.org/abs/2606.22873"}, "sourceUpdatedAt": "2026-06-25T18:44:01Z", "topics": ["Multimodal", "Reasoning"], "venueAttempts": [], "whyItMatters": "It provides a repeatable way to compare guardrails under both fixed safety tasks and runtime policy changes."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
