{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.08542", "kind": "paper", "url": "https://arxiv.org/abs/2608.08542"}], "authors": ["Yu Ma", "Hongli Shi", "Jing Li", "Xinran Xu", "Weiwei Hou"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.08542", "source_url": "https://arxiv.org/abs/2608.08542"}, "categories": ["General AI", "Robustness", "Safety", "Safety & Trustworthiness"], "description": {"en": "SkillSafe-Bench evaluates skill-merged LLMs on static refusal, adaptive jailbreak robustness, and capability retention using a two-judge AND rule, across multiple open-weight bases and attack types."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.08542", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.08542", "title": "SkillSafe-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["UNSW Sydney", "University of Technology Sydney", "Zhejiang University", "Australian National University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.08542", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.08542", "source_url": "https://openalex.org/W7202203323"}, "key": "claire-radar:2608.08542", "modality": null, "name": "SkillSafe-Bench", "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_skillsafe-bench_e266705a", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "bc00858e6328ccdff7f3bd0d3a070f5b4e5972e91bc389e20cca485c0aadaf45", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2608.08542"}, "publisher": null, "released": "2026-08-09", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.08542", "source_url": "https://arxiv.org/abs/2608.08542"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-08542", "source": "claire_radar", "source_benchmark_id": "2608.08542", "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": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2608.08542", "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", "state": "unavailable"}}, "status": "unavailable"}, "capabilities": ["Robustness"], "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": "cde271257299f16d1e2889b842f9edf7caa9647aa300443ef945774ed634650d", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "SkillSafe-Bench evaluates skill-merged LLMs on static refusal, adaptive jailbreak robustness, and capability retention using a two-judge AND rule, across multiple open-weight bases and attack types.", "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 SkillSafe-Bench, a controlled benchmark that scores skill-merged models on static refusal, adaptive jailbreak robustness, and capability retention under a conservative two-judge AND rule."}, "familyId": "bmf_1ffd6ecd5782", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_skillsafe-bench_e266705a", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2608.08542", "paper": "https://arxiv.org/abs/2608.08542", "pdf": "https://arxiv.org/pdf/2608.08542", "project": null, "report": "https://arxiv.org/abs/2608.08542"}, "metrics": [], "motivation": "Model merging has become the default way to give an aligned language model new skills without retraining: a practitioner folds task vectors from math, code, or domain specialists into a safety-aligned base using task arithmetic, TIES, or DARE.", "name": "SkillSafe-Bench", "oneLine": "SkillSafe-Bench evaluates skill-merged LLMs on static refusal, adaptive jailbreak robustness, and capability retention using a two-judge AND rule, across multiple open-weight bases and attack types.", "paperTitle": "When Skills Meet Safety: Benchmarking and Characterizing the Adaptive Jailbreak Robustness of Skill-Merged LLMs", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.08542", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {"90d": {"confidence": "Low", "coverage": 0.0, "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": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v13", "rank": null, "score": null}}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-08-09", "source": {"authors": ["Yu Ma", "Hongli Shi", "Jing Li", "Xinran Xu", "Weiwei Hou"], "categories": ["cs.LG"], "id": "2608.08542", "title": "When Skills Meet Safety: Benchmarking and Characterizing the Adaptive Jailbreak Robustness of Skill-Merged LLMs", "type": "arxiv", "url": "https://arxiv.org/abs/2608.08542"}, "sourceUpdatedAt": "2026-08-09T07:41:06Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "It exposes that static safety does not predict robustness to adaptive attacks, providing a more accurate safety evaluation for model merging and guiding safer merging practices."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
