{"record": {"aliases": ["SWORD: Wikidata-based Distortions Reveal Hidden Cross-Lingual Inconsistencies in LLM Factual Error Rejection"], "artifacts": [{"id": "arxiv:2609.09349", "kind": "paper", "url": "https://arxiv.org/abs/2609.09349"}], "authors": ["Sanghyeok Park", "Minji Kang", "Hosung Kwak", "Jinhyuk Yun"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.09349", "source_url": "https://arxiv.org/abs/2609.09349"}, "categories": ["Factuality", "General AI", "Language & Knowledge", "cs.CL"], "description": {"en": "A benchmark of syntactically correct but factually incorrect statements across eight languages for evaluating multilingual factual error rejection."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.09349", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.09349", "title": "SWORD"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Soongsil University", "Korea Advanced Institute of Science and Technology"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2609.09349", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2609.09349", "source_url": "https://openalex.org/W7212181657"}, "key": "claire-radar:2609.09349", "modality": null, "name": "SWORD", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "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_sword-wikidata-based-distortions-reveal-hi_9208f906", "origin_source": "arxiv", "recognition_confidence": "0.6", "record_sha256": "cc1f463e9e7444a94833f65517a9169dbbd0b93bc0af732936b40c052618366d", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-name-audit-deferred", "reviewed_at": "2026-09-10T06:21:31.238194Z", "source_url": "https://arxiv.org/abs/2609.09349"}, "publisher": null, "released": "2026-09-08", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.09349", "source_url": "https://arxiv.org/abs/2609.09349"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-09349", "source": "claire_radar", "source_benchmark_id": "2609.09349", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["SWORD: Wikidata-based Distortions Reveal Hidden Cross-Lingual Inconsistencies in LLM Factual Error Rejection"], "annotation": "Mixed", "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/2609.09349", "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"}, "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "Cross-lingual factual consistency is a growing concern, and the benchmark's multilingual focus may draw interest from NLP researchers.", "score": 45}, "capabilities": ["Factuality"], "confidence": "Medium", "construction": "Aggregate Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-10T06:21:31.238194Z", "inputHash": "ee10c20e307b685d28308727bdd94598cc990d35b553b40ce0c1bf8f14470f3c", "model": "deepseek-v4-pro"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We introduce Systematic Wikidata-based Object-Relation Distortion (SWORD), a benchmark that evaluates whether models consistently reject factual errors across languages.", "canonicalNameSource": "abstract", "decisionReason": "No public dataset or code link is provided, making the benchmark inaccessible for external evaluation.", "evaluationEvidence": {"reuse": "The abstract describes a distortion-based evaluation that surfaces insights obscured by conventional benchmarks, implying a reproducible evaluation protocol for multilingual factual reasoning.", "scoring": "The benchmark evaluates whether models consistently reject factual errors across languages, with performance measured by accuracy on distorted statements.", "task": "SWORD generates syntactically well-formed but factually incorrect statements in eight widely spoken languages through controlled perturbations of Wikidata triples, ranging from random entity substitutions to semantically plausible property-based selections."}, "model": "deepseek-v4-pro", "reviewedAt": "2026-09-10T06:21:31.238194Z", "state": "ai-name-audit-deferred"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A benchmark of syntactically correct but factually incorrect statements across eight languages for evaluating multilingual factual error rejection.", "displayEligible": false, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We introduce Systematic Wikidata-based Object-Relation Distortion (SWORD), a benchmark that evaluates whether models consistently reject factual errors across languages."}, "familyId": "bmf_ce70b3a84b42", "firstSeenAt": "2026-09-10", "heat": null, "id": "bm_sword-wikidata-based-distortions-reveal-hi_9208f906", "indexedAt": "2026-09-10T00:47:26.889471Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2609.09349", "pdf": "https://arxiv.org/pdf/2609.09349", "project": null, "report": "https://arxiv.org/abs/2609.09349"}, "metrics": [], "motivation": "Modern LLMs demonstrate impressive multilingual performance, yet standard benchmarks primarily reward selecting correct answers rather than evaluating genuine factual understanding.", "name": "SWORD", "oneLine": "A benchmark of syntactically correct but factually incorrect statements across eight languages for evaluating multilingual factual error rejection.", "paperTitle": "SWORD: Wikidata-based Distortions Reveal Hidden Cross-Lingual Inconsistencies in LLM Factual Error Rejection", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.09349", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-10T00:47:26.889471Z"}, "publications": [], "ranking": {"30d": {"confidence": "Low", "coverage": 0.0, "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": 30, "historyDate": "2026-09-05", "rank": null, "score": null, "windowDays": 30}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": null, "value": null}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v13", "rank": null, "score": null}, "90d": {"confidence": "Low", "coverage": 0.0, "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": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v13", "rank": null, "score": null}}, "readiness": "Paper only", "recognitionConfidence": 0.6, "relation": "introduces", "releasedAt": "2026-09-08", "source": {"authors": ["Sanghyeok Park", "Minji Kang", "Hosung Kwak", "Jinhyuk Yun"], "categories": ["cs.CL"], "id": "2609.09349", "title": "SWORD: Wikidata-based Distortions Reveal Hidden Cross-Lingual Inconsistencies in LLM Factual Error Rejection", "type": "arxiv", "url": "https://arxiv.org/abs/2609.09349"}, "sourceUpdatedAt": "2026-09-08T18:39:01Z", "topics": ["cs.CL"], "venueAttempts": [], "whyItMatters": "Exposes cross-lingual inconsistencies in LLM factual verification, beyond aggregate accuracy metrics."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
