{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2609.05117", "kind": "paper", "url": "https://arxiv.org/abs/2609.05117"}], "authors": ["Thu-Hien Trinh-Thi", "Hai-Yen Vong", "Thanh-Ha Ung-Dung", "Tram Ho"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.05117", "source_url": "https://arxiv.org/abs/2609.05117"}, "categories": ["General AI", "Safety", "Safety & Trustworthiness"], "description": {"en": "Evaluates LLM safety behaviors across four risk domains and four threat levels using a six-label behavior scale judged by two LLMs."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.05117", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.05117", "title": "TIER"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Vietnam National University Ho Chi Minh City"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2609.05117", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2609.05117", "source_url": "https://openalex.org/W7211905752"}, "key": "claire-radar:2609.05117", "modality": null, "name": "TIER", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "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_tier_4396e23d", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "7def284b4d19990e21fc2aa8f05550b77e8b42c8cd4acd009a56ae7c30c2947b", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-09T09:03:45.514878Z", "source_url": "https://arxiv.org/abs/2609.05117"}, "publisher": null, "released": "2026-09-04", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.05117", "source_url": "https://arxiv.org/abs/2609.05117"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-05117", "source": "claire_radar", "source_benchmark_id": "2609.05117", "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/2609.05117", "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"}, "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "The topic of safety benchmarks for LLMs attracts moderate interest, but the specific framing may limit broad appeal.", "score": 45}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-09T09:03:45.514878Z", "inputHash": "b2395b34153b7fa47adf1822791e7818e21d6feb4e374d980b63f01241f8c497", "model": "deepseek-v4-pro"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We introduce TIER, a Threat Implicitness Benchmark for behavioral safety evaluation of LLMs.", "canonicalNameSource": "abstract", "decisionReason": "The benchmark defines a repeatable evaluation with a clear scoring contract and has been applied to six open-weight models, demonstrating public reusability.", "evaluationEvidence": {"reuse": "Experiments on six open-weight LLMs show that safety behaviors evolve gradually across threat levels rather than shifting directly from refusal to compliance.", "scoring": "Responses are assessed using a six-label behavior scale and two independent LLM judges.", "task": "Responses are assessed using a six-label behavior scale and two independent LLM judges."}, "model": "deepseek-v4-pro", "reviewedAt": "2026-09-09T09:03:45.514878Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates LLM safety behaviors across four risk domains and four threat levels using a six-label behavior scale judged by two LLMs.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence"], "snippet": "We introduce TIER, a Threat Implicitness Benchmark for behavioral safety evaluation of LLMs."}, "familyId": "bmf_e29ccc00b2e5", "firstSeenAt": "2026-09-07", "heat": null, "id": "bm_tier_4396e23d", "indexedAt": "2026-09-07T06:22:02.605340Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2609.05117", "pdf": "https://arxiv.org/pdf/2609.05117", "project": null, "report": "https://arxiv.org/abs/2609.05117"}, "metrics": [], "motivation": "Current LLM safety benchmarks largely rely on binary metrics, overlooking how models respond to harmful prompts with varying threat implicitness.", "name": "TIER", "oneLine": "Evaluates LLM safety behaviors across four risk domains and four threat levels using a six-label behavior scale judged by two LLMs.", "paperTitle": "TIER: Threat Implicitness Benchmark for Evaluating LLM Safety Behaviors", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.05117", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-07T06:22:02.605340Z"}, "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": "2026-09-04", "source": {"authors": ["Thu-Hien Trinh-Thi", "Hai-Yen Vong", "Thanh-Ha Ung-Dung", "and Tram Ho"], "categories": ["cs.CR", "cs.AI"], "id": "2609.05117", "title": "TIER: Threat Implicitness Benchmark for Evaluating LLM Safety Behaviors", "type": "arxiv", "url": "https://arxiv.org/abs/2609.05117"}, "sourceUpdatedAt": "2026-09-04T13:13:11Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "Provides behavior-aware safety evaluation beyond binary metrics, revealing distinct response distributions even among models with similar attack success rates."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
