{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2607.23545", "kind": "paper", "url": "https://arxiv.org/abs/2607.23545"}, {"id": "gh:g1moon/language-shapes-ih", "kind": "repo", "url": "https://github.com/g1moon/Language-Shapes-IH"}], "authors": ["Jiwon Moon", "Yerin Hwang", "Kyomin Jung"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.23545", "source_url": "https://arxiv.org/abs/2607.23545"}, "categories": ["General AI", "Language & Knowledge", "cs.CL"], "description": {"en": "Evaluates instruction hierarchy compliance in multilingual LLMs using same-language and cross-language conflicts across six languages, four domains (rule-following, safety, task-execution, persona), and three hierarchy types (system-user, system-tool, user-tool)."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.23545", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.23545", "title": "XIH-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2607.23545", "modality": null, "name": "XIH-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_xih-bench_b0306814", "origin_source": "arxiv", "recognition_confidence": "0.95", "record_sha256": "603d05239ab11e7e855d9cbcc9ed813dd8277a9219e24cebff6007c7d50d1720", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2607.23545"}, "publisher": null, "released": "2026-07-26", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.23545", "source_url": "https://arxiv.org/abs/2607.23545"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-23545", "source": "claire_radar", "source_benchmark_id": "2607.23545", "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": "https://github.com/g1moon/Language-Shapes-IH", "githubScope": "benchmark_repo", "githubStars": 2, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 1, "hfPaperUrl": "https://huggingface.co/papers/2607.23545", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "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": "a550351d2dc3302ee2fca157c9db57278a19262f3fd0cf3b99c44e2e41ffcd8c", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates instruction hierarchy compliance in multilingual LLMs using same-language and cross-language conflicts across six languages, four domains (rule-following, safety, task-execution, persona), and three hierarchy types (system-user, system-tool, user-tool).", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "benchmark release stated in one sentence", "public artifact URL"], "snippet": "We introduce XIH-Bench, a benchmark for multilingual IH evaluation with both same-language and cross-language conflicts across six languages, four domains, and three IH settings."}, "familyId": "bmf_a796a6dc7c9f", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [{"evidence": [{"basis": "existing-code-link", "linkedUrl": "https://github.com/g1moon/Language-Shapes-IH", "url": "https://github.com/g1moon/Language-Shapes-IH"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/g1moon/Language-Shapes-IH", "url": "https://huggingface.co/papers/2607.23545"}, {"basis": "linked-from-report", "linkedUrl": "https://github.com/g1moon/Language-Shapes-IH", "url": "https://arxiv.org/abs/2607.23545"}], "scope": "project_repo", "url": "https://github.com/g1moon/Language-Shapes-IH", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_xih-bench_b0306814", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/g1moon/Language-Shapes-IH", "data": null, "hfPaper": "https://huggingface.co/papers/2607.23545", "paper": "https://arxiv.org/abs/2607.23545", "pdf": "https://arxiv.org/pdf/2607.23545", "project": null, "report": "https://arxiv.org/abs/2607.23545"}, "metrics": [], "motivation": "Instruction hierarchy (IH) requires models to prioritize instructions by source, ensuring that higher-priority instructions override lower-priority ones.", "name": "XIH-Bench", "oneLine": "Evaluates instruction hierarchy compliance in multilingual LLMs using same-language and cross-language conflicts across six languages, four domains (rule-following, safety, task-execution, persona), and three hierarchy types (system-user, system-tool, user-tool).", "paperTitle": "Language Shapes Instruction Hierarchy Compliance in Multilingual LLMs", "primaryDomain": "General AI", "publication": {"evidence": "Accepted to EMNLP 2026 (Main). Code and data are available at https://github.com/g1moon/Language-Shapes-IH", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2607.23545", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "EMNLP 2026 (Main)", "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "g1moon", "organizationType": "community", "role": "benchmark-publisher", "sourceUrl": "https://github.com/g1moon/Language-Shapes-IH"}], "ranking": {"90d": {"confidence": "Medium", "coverage": 0.7, "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": 0.1466449917496476, "value": 2}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.12247537038455211, "value": 1}}, "confidence": "Medium", "coverage": 0.7, "rank": 371, "score": 25}, "method": "attention-ranking-v13", "rank": 371, "score": 25}}, "readiness": "Runnable", "recognitionConfidence": 0.95, "relation": "introduces", "releasedAt": "2026-07-26", "source": {"authors": ["Jiwon Moon", "Yerin Hwang", "Kyomin Jung"], "categories": ["cs.CL"], "id": "2607.23545", "title": "Language Shapes Instruction Hierarchy Compliance in Multilingual LLMs", "type": "arxiv", "url": "https://arxiv.org/abs/2607.23545"}, "sourceUpdatedAt": "2026-07-26T08:50:28Z", "topics": ["cs.CL"], "venueAttempts": [{"decisionRaw": "Accepted to EMNLP 2026 (Main). Code and data are available at https://github.com/g1moon/Language-Shapes-IH", "evidence": [{"level": "author-claim", "observedAt": "2026-10-14T01:46:58.058199Z", "rawValue": "Accepted to EMNLP 2026 (Main). Code and data are available at https://github.com/g1moon/Language-Shapes-IH", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2607.23545"}], "reviewStatus": "accepted", "venueName": "EMNLP 2026 (Main)"}], "whyItMatters": "Existing instruction hierarchy benchmarks are largely English-centric, leaving a gap in assessing multilingual safety and reliability. This benchmark provides a reusable protocol to measure how language choice affects model compliance, supporting safer deployment in multilingual contexts."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
