{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2609.12808", "kind": "paper", "url": "https://arxiv.org/abs/2609.12808"}], "authors": ["Guangsheng Yu", "Yanna Jiang", "Qin Wang", "Baihe Ma", "Xu Wang"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.12808", "source_url": "https://arxiv.org/abs/2609.12808"}, "categories": ["General AI", "Language & Knowledge", "cs.AI"], "description": {"en": "K-Bench is described as a benchmark that scores LLM unlearning under agentic deployment."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.12808", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.12808", "title": "K-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["University of Technology Sydney", "Commonwealth Scientific and Industrial Research Organisation"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2609.12808", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2609.12808", "source_url": "https://openalex.org/W7212904611"}, "key": "claire-radar:2609.12808", "modality": null, "name": "K-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": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "false", "origin_record_id": "bm_k-bench_92156d34", "origin_source": "arxiv", "recognition_confidence": "0.85", "record_sha256": "2c4ada303f339a67bfbe939eb538be93c24208af739e71cfdcb808de78ab6b61", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-name-audit-deferred", "reviewed_at": "2026-09-14T10:05:20.149847Z", "source_url": "https://arxiv.org/abs/2609.12808"}, "publisher": null, "released": "2026-09-11", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.12808", "source_url": "https://arxiv.org/abs/2609.12808"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-12808", "source": "claire_radar", "source_benchmark_id": "2609.12808", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Mixed", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "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": 0, "hfPaperUrl": "https://huggingface.co/papers/2609.12808", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "No released artifacts, task inputs, or public path are supplied and the description is minimal, limiting expected first-week attention.", "score": 0}, "capabilities": [], "confidence": "Medium", "construction": "Aggregate Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-15T09:37:10.086794Z", "inputHash": "bf3f8dc5b8571bed8ee8cced3346a52f4a2df11736feb96d214998e1a8cb0f4b", "model": "deepseek-v4-flash"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We introduce K-Bench, a benchmark that scores LLM unlearning under agentic deployment.", "canonicalNameSource": "abstract", "decisionReason": "Insufficient evidence of a formally named benchmark with a repeatable evaluation object, comparable scoring contract, and public path for other teams to run or inspect it.", "evaluationEvidence": {"reuse": "On structured retrieval, the secret stays verbatim in the tool-observation channel and the aggregate leak rate is unchanged.", "scoring": "The K-Score is computed separately for each source and credits forgetting only when the agent remains usable.", "task": "K-Bench inspects all six channels a ReAct agent exposes, including its chain-of-thought (CoT), tool calls and tool observations, and elicited summary."}, "model": "deepseek-v4-pro", "reviewedAt": "2026-09-14T10:05:20.149847Z", "state": "ai-name-audit-deferred"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "K-Bench is described as a benchmark that scores LLM unlearning under agentic deployment.", "displayEligible": false, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence"], "snippet": "We introduce K-Bench, a benchmark that scores LLM unlearning under agentic deployment."}, "familyId": "bmf_5a350d00e40c", "firstSeenAt": "2026-09-14", "heat": null, "id": "bm_k-bench_92156d34", "indexedAt": "2026-09-14T06:39:43.852630Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2609.12808", "pdf": "https://arxiv.org/pdf/2609.12808", "project": null, "report": "https://arxiv.org/abs/2609.12808"}, "metrics": [], "motivation": "Unlearning benchmarks such as TOFU and MUSE certify forgetting by reading the model's final answer, where a model that refuses to answer already counts as having forgotten.", "name": "K-Bench", "oneLine": "K-Bench is described as a benchmark that scores LLM unlearning under agentic deployment.", "paperTitle": "K-Bench: A Benchmark for LLM Unlearning in Agentic Deployments", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.12808", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-14T06:39:43.852630Z"}, "publications": [], "ranking": {"30d": {"confidence": "Low", "coverage": 0.3, "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": 30, "historyDate": "2026-09-05", "rank": null, "score": null, "windowDays": 30}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": null, "value": null}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.0, "value": 0}}, "confidence": "Low", "coverage": 0.3, "rank": null, "score": 35}, "method": "attention-ranking-v13", "rank": null, "score": 35}, "90d": {"confidence": "Low", "coverage": 0.15, "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": 0.0, "value": 0}}, "confidence": "Low", "coverage": 0.15, "rank": null, "score": 43}, "method": "attention-ranking-v13", "rank": null, "score": 43}}, "readiness": "Paper only", "recognitionConfidence": 0.85, "relation": "introduces", "releasedAt": "2026-09-11", "source": {"authors": ["Guangsheng Yu and Yanna Jiang and Qin Wang and Baihe Ma and Xu Wang"], "categories": ["cs.AI"], "id": "2609.12808", "title": "K-Bench: A Benchmark for LLM Unlearning in Agentic Deployments", "type": "arxiv", "url": "https://arxiv.org/abs/2609.12808"}, "sourceUpdatedAt": "2026-09-11T13:04:13Z", "topics": ["cs.AI"], "venueAttempts": [], "whyItMatters": "Unlearning evaluation largely focuses on static prompting, while agentic deployments introduce tool use and multi-step behavior; a benchmark targeting that setting could inform safety and compliance assessment."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
