{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2607.05910", "kind": "paper", "url": "https://arxiv.org/abs/2607.05910"}, {"id": "gh:ssmisya/policyshiftguard", "kind": "repo", "url": "https://github.com/ssmisya/PolicyShiftGuard"}], "authors": ["Mingyang Song", "Luxin Xu", "Haoyu Sun", "Minzhou Pan", "Yu Cheng", "Bo Li"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.05910", "source_url": "https://arxiv.org/abs/2607.05910"}, "categories": ["General AI", "Safety", "Safety & Trustworthiness"], "description": {"en": "PolicyShiftBench evaluates policy-adaptive image guardrailing: given an image and a current policy, a model must output a pass/block decision plus optional violated category IDs. The benchmark comprises 2,000 policy-discriminative instances over 265 images, each paired with multiple policy-conditioned prompts. Scoring uses binary pass/block accuracy and category attribution metrics."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.05910", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.05910", "title": "PolicyShiftBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Fudan University", "Tongji University", "Chinese University of Hong Kong"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2607.05910", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2607.05910", "source_url": "https://openalex.org/W7167679369"}, "key": "claire-radar:2607.05910", "modality": null, "name": "PolicyShiftBench", "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_policyshiftbench_8ca9566e", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "70f35864dd9fae0f5a8b4034c07a68973aa10222549043460328ef368245dfb6", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2607.05910"}, "publisher": null, "released": "2026-07-07", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.05910", "source_url": "https://arxiv.org/abs/2607.05910"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-05910", "source": "claire_radar", "source_benchmark_id": "2607.05910", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/ssmisya/PolicyShiftGuard", "githubScope": "benchmark_repo", "githubStars": 22, "hfDailySubmittedAt": "2026-07-16T00:00:00.000Z", "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 30, "hfPaperUrl": "https://huggingface.co/papers/2607.05910", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}}, "status": "fresh"}, "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": "96af7e72f658f21967c3d891f085a7c6c3e49392ed4692678485a7b50b0f5cd2", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "PolicyShiftBench evaluates policy-adaptive image guardrailing: given an image and a current policy, a model must output a pass/block decision plus optional violated category IDs. The benchmark comprises 2,000 policy-discriminative instances over 265 images, each paired with multiple policy-conditioned prompts. Scoring uses binary pass/block accuracy and category attribution metrics.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We introduce PolicyShiftBench, a comprehensive benchmark with 2,000 policy-discriminative instances over 265 images, where each image is paired with 7.55 policy-conditioned prompts on average to test whether models adapt to the active policy rather than relying on image-level safety priors."}, "familyId": "bmf_1a011cbd85dd", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [{"evidence": [{"basis": "existing-code-link", "linkedUrl": "https://github.com/ssmisya/PolicyShiftGuard", "url": "https://github.com/ssmisya/PolicyShiftGuard"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/ssmisya/PolicyShiftGuard", "url": "https://huggingface.co/papers/2607.05910"}], "scope": "project_repo", "url": "https://github.com/ssmisya/PolicyShiftGuard", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_policyshiftbench_8ca9566e", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/ssmisya/PolicyShiftGuard", "data": null, "hfPaper": "https://huggingface.co/papers/2607.05910", "paper": "https://arxiv.org/abs/2607.05910", "pdf": "https://arxiv.org/pdf/2607.05910", "project": null, "report": "https://arxiv.org/abs/2607.05910"}, "metrics": [], "motivation": "Image guardrails are typically trained and evaluated under a fixed safety policy, implicitly treating safety as an intrinsic property of an image.", "name": "PolicyShiftBench", "oneLine": "PolicyShiftBench evaluates policy-adaptive image guardrailing: given an image and a current policy, a model must output a pass/block decision plus optional violated category IDs. The benchmark comprises 2,000 policy-discriminative instances over 265 images, each paired with multiple policy-conditioned prompts. Scoring uses binary pass/block accuracy and category attribution metrics.", "paperTitle": "PolicyShiftGuard: Benchmarking and Improving Policy-Adaptive Image Guardrails", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.05910", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {"90d": {"confidence": "Medium", "coverage": 0.7, "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": 0.418532113802802, "value": 22}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.606767028072414, "value": 30}}, "confidence": "Medium", "coverage": 0.7, "rank": 91, "score": 47}, "method": "attention-ranking-v13", "rank": 91, "score": 47}}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-07-07", "source": {"authors": ["Mingyang Song", "Luxin Xu", "Haoyu Sun", "Minzhou Pan", "Yu Cheng", "Bo Li"], "categories": ["cs.CV", "cs.AI", "cs.CL"], "id": "2607.05910", "title": "PolicyShiftGuard: Benchmarking and Improving Policy-Adaptive Image Guardrails", "type": "arxiv", "url": "https://arxiv.org/abs/2607.05910"}, "sourceUpdatedAt": "2026-07-07T07:04:28Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "Existing image safety benchmarks assume safety is a fixed property of an image, whereas real deployments vary policies across products and regions. PolicyShiftBench measures whether models can bind image evidence to the active policy rather than relying on image-level priors, providing a practical evaluation for guardrails in dynamic policy environments."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
