{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.10258", "kind": "paper", "url": "https://arxiv.org/abs/2608.10258"}], "authors": ["Waleed Jamil", "Raphael Schmitt"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.10258", "source_url": "https://arxiv.org/abs/2608.10258"}, "categories": ["Biology & Drug Discovery", "Safety", "Safety & Trustworthiness"], "description": {"en": "TAF-MED is a physician-reviewed benchmark of 500 fixed three-turn medical safety scenarios, evaluating LLM responses for unsafe guidance across multi-turn dialogues."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.10258", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.10258", "title": "TAF-MED"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Oldham Council", "University of Freiburg", "University Medical Center Freiburg", "Technical University of Munich"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.10258", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.10258", "source_url": "https://openalex.org/W7202279193"}, "key": "claire-radar:2608.10258", "modality": null, "name": "TAF-MED", "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_taf-med_ece2193a", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "c970d80ca9b65bdfd3879bdfb36e07203eba25b3130706ac6f06edb1d5978133", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2608.10258"}, "publisher": null, "released": "2026-08-10", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.10258", "source_url": "https://arxiv.org/abs/2608.10258"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-10258", "source": "claire_radar", "source_benchmark_id": "2608.10258", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["Biology & Drug Discovery"], "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": 0, "hfPaperUrl": "https://huggingface.co/papers/2608.10258", "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"}, "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": "c67df0ae4961ae1cee8f27b4ea1848b457e8ef6298a8975347d93513da9dab4a", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "TAF-MED is a physician-reviewed benchmark of 500 fixed three-turn medical safety scenarios, evaluating LLM responses for unsafe guidance across multi-turn dialogues.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence"], "snippet": "We introduce TAF-MED, a physician-reviewed benchmark of 500 fixed three-turn scenarios, and evaluate eight LLMs across 4,000 conversations."}, "familyId": "bmf_437ad9931b00", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_taf-med_ece2193a", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Pharma & Biotech"], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2608.10258", "paper": "https://arxiv.org/abs/2608.10258", "pdf": "https://arxiv.org/pdf/2608.10258", "project": null, "report": "https://arxiv.org/abs/2608.10258"}, "metrics": [], "motivation": "Large language models (LLMs) increasingly provide conversational health information that may influence treatment decisions, yet existing benchmarks do not isolate whether medication-safety boundaries persist across follow-ups after explicit self-treatment intent.", "name": "TAF-MED", "oneLine": "TAF-MED is a physician-reviewed benchmark of 500 fixed three-turn medical safety scenarios, evaluating LLM responses for unsafe guidance across multi-turn dialogues.", "paperTitle": "TAF-MED: Multi-Turn Safety Refusal Collapse in LLMs Under Declared Self-Treatment Intent", "primaryDomain": "Biology & Drug Discovery", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.10258", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {"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": 1.0, "relation": "introduces", "releasedAt": "2026-08-10", "source": {"authors": ["Waleed Jamil", "Raphael Schmitt"], "categories": ["cs.CL", "cs.AI"], "id": "2608.10258", "title": "TAF-MED: Multi-Turn Safety Refusal Collapse in LLMs Under Declared Self-Treatment Intent", "type": "arxiv", "url": "https://arxiv.org/abs/2608.10258"}, "sourceUpdatedAt": "2026-08-10T21:43:58Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "Addresses the evaluation gap where first-turn safety is an incomplete proxy for conversational safety persistence, providing a protocol for assessing model behavior across complete dialogue trajectories in medical contexts."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
