{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2605.27649", "kind": "paper", "url": "https://arxiv.org/abs/2605.27649"}], "authors": ["Qishi Zhan", "Minxuan Hu", "Seoyeon Jang", "Lei Zhao", "Ziheng Chen", "Man Liang", "Xinyue Xiang", "Jiaxin Liu", "Guansu Wang", "Liang He"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2605.27649", "source_url": "https://arxiv.org/abs/2605.27649"}, "categories": ["General AI", "Language & Knowledge", "cs.CL"], "description": {"en": "MTM-Bench is a controlled benchmark for language-conditioned task execution in multilingual settings, enumerating all 27 instruction-content-response language triplets across English, Spanish, and Chinese. It contains 2,430 instances per model across semantic reversal, final-state extraction, and language purity tasks, with decomposed metrics for semantic correctness, language adherence, constraint satisfaction, and joint success."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2605.27649", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2605.27649", "title": "MTM-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["University of Pennsylvania", "Cornell University", "University of Illinois Urbana-Champaign"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2605.27649", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2605.27649", "source_url": "https://openalex.org/W7162684862"}, "key": "claire-radar:2605.27649", "modality": null, "name": "MTM-Bench", "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_mtm-bench_e1a4ed71", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "d7936be3ea7c5112f3d0804f5617789a8a706c8f1bb037b0b450a293aad92816", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2605.27649"}, "publisher": null, "released": "2026-05-26", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2605.27649", "source_url": "https://arxiv.org/abs/2605.27649"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2605-27649", "source": "claire_radar", "source_benchmark_id": "2605.27649", "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/2605.27649", "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": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "753bd38a7699932ff39b720a13aa24419a9b3088915cdaae11a259e0b053709b", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "MTM-Bench is a controlled benchmark for language-conditioned task execution in multilingual settings, enumerating all 27 instruction-content-response language triplets across English, Spanish, and Chinese. It contains 2,430 instances per model across semantic reversal, final-state extraction, and language purity tasks, with decomposed metrics for semantic correctness, language adherence, constraint satisfaction, and joint success.", "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": "We introduce MTM-Bench, a controlled benchmark for language-conditioned task execution in which each instance is defined by a triplet \\((L_{\\text{instr}}, L_{\\text{content}}, L_{\\text{resp}})\\)."}, "familyId": "bmf_a0d4dfc5a508", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_mtm-bench_e1a4ed71", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2605.27649", "paper": "https://arxiv.org/abs/2605.27649", "pdf": "https://arxiv.org/pdf/2605.27649", "project": null, "report": "https://arxiv.org/abs/2605.27649"}, "metrics": [], "motivation": "Multilingual LLMs are increasingly used when instruction, source content, and required response languages do not coincide.", "name": "MTM-Bench", "oneLine": "MTM-Bench is a controlled benchmark for language-conditioned task execution in multilingual settings, enumerating all 27 instruction-content-response language triplets across English, Spanish, and Chinese. It contains 2,430 instances per model across semantic reversal, final-state extraction, and language purity tasks, with decomposed metrics for semantic correctness, language adherence, constraint satisfaction, and joint success.", "paperTitle": "Disentangling Language Roles in Multilingual LLM Task Execution", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2605.27649", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-05-26", "source": {"authors": ["Qishi Zhan", "Minxuan Hu", "Seoyeon Jang", "Lei Zhao", "Ziheng Chen", "Man Liang", "Xinyue Xiang", "Jiaxin Liu", "Guansu Wang", "Liang He"], "categories": ["cs.CL", "cs.LG"], "id": "2605.27649", "title": "Disentangling Language Roles in Multilingual LLM Task Execution", "type": "arxiv", "url": "https://arxiv.org/abs/2605.27649"}, "sourceUpdatedAt": "2026-05-26T20:09:34Z", "topics": ["cs.CL"], "venueAttempts": [], "whyItMatters": "Multilingual LLMs are used when instruction, source content, and response languages differ, yet existing evaluations rarely isolate these roles. MTM-Bench provides a fully crossed design to attribute degradation to specific language roles, revealing that response-slot mismatch drives most performance loss and that mismatch count is not a monotonic predictor of difficulty."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
