{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.23375", "kind": "paper", "url": "https://arxiv.org/abs/2606.23375"}], "authors": ["Arthur Wuhrmann", "Gaetan Stein", "Daniel Brunner", "Andrei Kucharavy"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.23375", "source_url": "https://arxiv.org/abs/2606.23375"}, "categories": ["General AI", "Language & Knowledge", "Safety"], "description": {"en": "TF-RefusalBench is a multilingual benchmark for criminal-law translation and summarization derived from public Swiss Supreme Court rulings. It contains 5,200 prompts across French, German, Italian, and English, designed to trigger refusals in LLMs."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.23375", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.23375", "title": "TF-RefusalBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Federal Supreme Court of Switzerland", "HES-SO Valais-Wallis"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.23375", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.23375", "source_url": "https://openalex.org/W7165626386"}, "key": "claire-radar:2606.23375", "modality": null, "name": "TF-RefusalBench", "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_tf-refusalbench_fb205626", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "305c6e816afb6740204e1caf64b065b1b8058f6d37141e495c7aa93be37ac7fa", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.23375"}, "publisher": null, "released": "2026-06-22", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.23375", "source_url": "https://arxiv.org/abs/2606.23375"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-23375", "source": "claire_radar", "source_benchmark_id": "2606.23375", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "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": null, "hfPaperUrl": "https://huggingface.co/papers/2606.23375", "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": [], "confidence": "High", "construction": "Transform Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "437f0e720580a98c5e3c2868201b51df55e7b88ae14ddf635fca01c055777d3c", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "TF-RefusalBench is a multilingual benchmark for criminal-law translation and summarization derived from public Swiss Supreme Court rulings. It contains 5,200 prompts across French, German, Italian, and English, designed to trigger refusals in LLMs.", "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": "To measure this phenomenon, we introduce TF-RefusalBench, a multilingual benchmark for criminal-law translation and summarization derived from public Swiss Supreme Court rulings."}, "familyId": "bmf_c969544e9b3f", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_tf-refusalbench_fb205626", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2606.23375", "paper": "https://arxiv.org/abs/2606.23375", "pdf": "https://arxiv.org/pdf/2606.23375", "project": null, "report": "https://arxiv.org/abs/2606.23375"}, "metrics": [], "motivation": "While the wider applicability of LLMs in the legal field is currently debated due to their reliability and the gravity of any errors, narrow uses with well-understood and mitigated risks have emerged.", "name": "TF-RefusalBench", "oneLine": "TF-RefusalBench is a multilingual benchmark for criminal-law translation and summarization derived from public Swiss Supreme Court rulings. It contains 5,200 prompts across French, German, Italian, and English, designed to trigger refusals in LLMs.", "paperTitle": "Measuring & Mitigating Over-Alignment for LLMs in Multilingual Criminal Law Courts", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.23375", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "arXiv", "organizationType": "community", "role": "benchmark-publisher", "sourceUrl": "https://arxiv.org/abs/2606.23375"}], "ranking": {}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-22", "source": {"authors": ["Arthur Wuhrmann", "Gaetan Stein", "Daniel Brunner", "Andrei Kucharavy"], "categories": ["cs.CL", "cs.AI"], "id": "2606.23375", "title": "Measuring & Mitigating Over-Alignment for LLMs in Multilingual Criminal Law Courts", "type": "arxiv", "url": "https://arxiv.org/abs/2606.23375"}, "sourceUpdatedAt": "2026-06-30T12:22:05Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "The benchmark addresses the challenge of evaluating over-alignment in LLMs performing legitimate legal tasks. It provides a standardized way to measure refusal behavior across languages and task types, aiding in the selection and mitigation of models for sensitive translation and summarization work."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
