{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.01666", "kind": "paper", "url": "https://arxiv.org/abs/2608.01666"}], "authors": ["Fengxian Ji", "Yuke Li", "Jingpu Yang", "Juanfan Wu", "Fan Zhang", "Zhexuan Cui", "Yu Xie", "Min Peng", "Qianqian Xie", "Xiuying Chen", "Zhuohan Xie"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.01666", "source_url": "https://arxiv.org/abs/2608.01666"}, "categories": ["General AI", "Information retrieval", "Safety & Trustworthiness", "cs.CL"], "description": {"en": "SciStyleBench diagnoses stylistic bias in LLM-based idea evaluation through controlled stylistic perturbations, metrics, and a mitigation extractor."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.01666", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.01666", "title": "SciStyleBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Wuhan University", "Mohamed bin Zayed University of Artificial Intelligence", "Nanyang Technological University", "University of Hong Kong"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.01666", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.01666", "source_url": "https://openalex.org/W7172431984"}, "key": "claire-radar:2608.01666", "modality": null, "name": "SciStyleBench", "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_scistylebench_232be1cc", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "5ab45dc8646ee9234e955f637b748d014fb22166ac45fe714e75ebcd630ea07d", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2608.01666"}, "publisher": null, "released": "2026-08-03", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.01666", "source_url": "https://arxiv.org/abs/2608.01666"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-01666", "source": "claire_radar", "source_benchmark_id": "2608.01666", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "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/2608.01666", "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": ["Information retrieval"], "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": "c5fa9ae335618287682507905754382d6f7077a6f33702e05d89bf2c1e48151a", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "SciStyleBench diagnoses stylistic bias in LLM-based idea evaluation through controlled stylistic perturbations, metrics, and a mitigation extractor.", "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 address this question, we propose SciStyleBench, a unified three-component benchmark for diagnosing and mitigating stylistic bias in LLM-based idea evaluation: (i) First, SciStyleStage, a three-stage evaluation environment that applies controlled stylistic perturbations to fixed scientific content across three settings no context, fixed-domain context, and open-domain retrieval context, covering 600 scientific ideas and 15 style variants, with 9,000 evaluation instances per setting; (ii) Seco"}, "familyId": "bmf_087086a27a87", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_scistylebench_232be1cc", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2608.01666", "paper": "https://arxiv.org/abs/2608.01666", "pdf": "https://arxiv.org/pdf/2608.01666", "project": null, "report": "https://arxiv.org/abs/2608.01666"}, "metrics": [], "motivation": "However, whether these judges truly evaluate the scientific substance of ideas or are influenced by superficial stylistic presentation remains an open question.", "name": "SciStyleBench", "oneLine": "SciStyleBench diagnoses stylistic bias in LLM-based idea evaluation through controlled stylistic perturbations, metrics, and a mitigation extractor.", "paperTitle": "Style Wins, Substance Loses: A Diagnosis of LLM-as-Judge in Idea Generation", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.01666", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {"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": 1.0, "relation": "introduces", "releasedAt": "2026-08-03", "source": {"authors": ["Fengxian Ji", "Yuke Li", "Jingpu Yang", "Juanfan Wu", "Fan Zhang", "Zhexuan Cui", "Yu Xie", "Min Peng", "Qianqian Xie", "Xiuying Chen", "Zhuohan Xie"], "categories": ["cs.CL", "cs.AI"], "id": "2608.01666", "title": "Style Wins, Substance Loses: A Diagnosis of LLM-as-Judge in Idea Generation", "type": "arxiv", "url": "https://arxiv.org/abs/2608.01666"}, "sourceUpdatedAt": "2026-08-04T03:10:26Z", "topics": ["cs.CL"], "venueAttempts": [], "whyItMatters": "Highlights the importance of style invariance in scientific idea evaluation and provides metrics and a mitigation module for improving LLM judges."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
