{"record": {"aliases": ["Register Shifts Break LLM Safety: A Bengali Benchmark with Culturally Grounded Harms"], "artifacts": [{"id": "arxiv:2608.22335", "kind": "paper", "url": "https://arxiv.org/abs/2608.22335v1"}], "authors": ["Naymul Islam", "Nusrat Jahan Lia", "Shubhashis Roy Dipta", "Sabik Bin Sultan", "Abdullah Khan Zehady"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.22335", "source_url": "https://arxiv.org/abs/2608.22335"}, "categories": ["General AI", "Safety", "Safety & Trustworthiness"], "description": {"en": "Evaluates LLM safety on 879 Bengali prompts spanning 17 harm categories and five prompting conditions varying language, writing style, and authority framing."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.22335v1", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.22335v1", "title": "BanglaSafe"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2608.22335", "modality": null, "name": "BanglaSafe", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"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_register-shifts-break-llm-safety-a-bengali_9b6dd4e8", "origin_source": "arxiv", "recognition_confidence": "0.45", "record_sha256": "1e80c88c3c23eef42178d205fa65bfcaebf689ef0b48cd8ced8277bae3404ea9", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-name-audit-deferred", "reviewed_at": "2026-08-27T04:18:56.318436Z", "source_url": "https://arxiv.org/abs/2608.22335v1"}, "publisher": null, "released": "2026-08-23", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.22335", "source_url": "https://arxiv.org/abs/2608.22335v1"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-22335", "source": "claire_radar", "source_benchmark_id": "2608.22335", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Register Shifts Break LLM Safety: A Bengali Benchmark with Culturally Grounded Harms"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-08-26", "attemptedAt": "2026-08-25T23:36:26.005677Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2608.22335", "observedAt": "2026-08-24T11:31:30.653116Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-08-25T23:36:26.005677Z", "state": "unavailable"}}, "status": "unavailable"}, "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "The benchmark targets a significant language and reports striking findings, but lack of public artifacts may limit initial engagement.", "score": 35}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-27T04:18:56.318436Z", "inputHash": "b1e237a6b94a51d418979ab59f8aac6faf2372c75b152fece31d7f160d9cd39e", "model": "deepseek-v4-pro"}, "curation": {"canonicalNameEvidence": "We introduce BanglaSafe, a benchmark of 879 Bengali prompts", "canonicalNameSource": "abstract", "decisionReason": "The benchmark is introduced in the paper and includes scored evaluations, but no public code, dataset, or project URL is provided, so the public reuse path is unclear.", "model": "deepseek-v4-pro", "reviewedAt": "2026-08-27T04:18:56.318436Z", "state": "ai-name-audit-deferred"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates LLM safety on 879 Bengali prompts spanning 17 harm categories and five prompting conditions varying language, writing style, and authority framing.", "displayEligible": false, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence"], "snippet": "We introduce BanglaSafe, a benchmark of 879 Bengali prompts combining 309 natively authored prompts with 570 expert-reviewed prompts, spanning 17 culturally grounded harm categories and five prompting conditions that vary language, writing style, and authority framing."}, "familyId": "bmf_e32f99ce3fcc", "firstSeenAt": "2026-08-25", "heat": null, "id": "bm_register-shifts-break-llm-safety-a-bengali_9b6dd4e8", "indexedAt": "2026-08-25T17:37:25.889905Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "http://arxiv.org/abs/2608.22335v1", "pdf": "https://arxiv.org/pdf/2608.22335v1", "project": null, "report": "http://arxiv.org/abs/2608.22335v1"}, "metrics": [], "motivation": "Bengali is the seventh-most-spoken language globally, yet LLM safety evaluation remains overwhelmingly English-centric.", "name": "BanglaSafe", "oneLine": "Evaluates LLM safety on 879 Bengali prompts spanning 17 harm categories and five prompting conditions varying language, writing style, and authority framing.", "paperTitle": "Register Shifts Break LLM Safety: A Bengali Benchmark with Culturally Grounded Harms", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "http://arxiv.org/abs/2608.22335v1", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-25T17:37:25.889905Z"}, "publications": [], "ranking": {"30d": {"confidence": "Low", "coverage": 0.0, "growth": {"components": {"githubStars": {"percentile": null, "value": null}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 30}, "level": {"components": {"githubStars": {"percentile": null, "value": null}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v11", "rank": null, "score": null}, "90d": {"confidence": "Low", "coverage": 0.0, "growth": {"components": {"githubStars": {"percentile": null, "value": null}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"percentile": null, "value": null}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v11", "rank": null, "score": null}}, "readiness": "Paper only", "recognitionConfidence": 0.45, "relation": "introduces", "releasedAt": "2026-08-23", "source": {"authors": ["Naymul Islam", "Nusrat Jahan Lia", "Shubhashis Roy Dipta", "Sabik Bin Sultan", "Abdullah Khan Zehady"], "categories": ["cs.CL", "cs.AI"], "id": "2608.22335", "title": "Register Shifts Break LLM Safety: A Bengali Benchmark with Culturally Grounded Harms", "type": "arxiv", "url": "http://arxiv.org/abs/2608.22335v1"}, "sourceUpdatedAt": "2026-08-23T10:09:00Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "Provides culturally grounded safety evaluation for Bengali, a widely spoken but underrepresented language in LLM safety testing, and reveals sensitivity to writing style."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
