{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.20527", "kind": "paper", "url": "https://arxiv.org/abs/2606.20527"}, {"id": "gh:timo-cavelius/stylisticbias", "kind": "repo", "url": "https://github.com/timo-cavelius/StylisticBias"}, {"kind": "website", "url": "https://hf.co/datasets/shaghayegh/stylistic-bias-dataset"}], "authors": ["Shaghayegh Kolli", "Timo Cavelius", "Nafiseh Nikeghbal", "Samantha Dalal", "Jana Diesner"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.20527", "source_url": "https://arxiv.org/abs/2606.20527"}, "categories": ["General AI", "Safety & Trustworthiness", "cs.CL"], "description": {"en": "StylisticBias is a benchmark of 25,000 images with controlled single-attribute variations to evaluate attribute-level social bias in multimodal LLMs. It measures how visual cues shift model judgments in 25 scenarios."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.20527", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.20527", "title": "StylisticBias"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Munich Center for Machine Learning", "Technical University of Munich", "Center for Information Technology", "Center For Policy Research"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.20527", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.20527", "source_url": "https://openalex.org/W7165376334"}, "key": "claire-radar:2606.20527", "modality": null, "name": "StylisticBias", "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_stylisticbias_ee73496d", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "14fa63f09e1dfa515c24ec11e40b4b60e9d84d0ed4278acf0038802139e918cc", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.20527"}, "publisher": null, "released": "2026-06-18", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.20527", "source_url": "https://arxiv.org/abs/2606.20527"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-20527", "source": "claire_radar", "source_benchmark_id": "2606.20527", "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": "https://github.com/timo-cavelius/StylisticBias", "githubScope": "benchmark_repo", "githubStars": 2, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 4, "hfPaperUrl": "https://huggingface.co/papers/2606.20527", "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": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-04", "state": "stale"}}, "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": "ff40a8290eef1406bddf205706c66c56314b768d849515374ac802451f91aad1", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "StylisticBias is a benchmark of 25,000 images with controlled single-attribute variations to evaluate attribute-level social bias in multimodal LLMs. It measures how visual cues shift model judgments in 25 scenarios.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence", "public artifact URL"], "snippet": "We introduce StylisticBias, a controlled benchmark for evaluating attribute-level social bias in MLLMs."}, "familyId": "bmf_110092e1dedb", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 1, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [{"evidence": [{"basis": "existing-code-link", "linkedUrl": "https://github.com/timo-cavelius/StylisticBias", "url": "https://github.com/timo-cavelius/StylisticBias"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/timo-cavelius/StylisticBias", "url": "https://huggingface.co/papers/2606.20527"}, {"basis": "linked-from-project", "linkedUrl": "https://github.com/timo-cavelius/StylisticBias", "url": "https://hf.co/datasets/shaghayegh/stylistic-bias-dataset"}, {"basis": "linked-from-report", "linkedUrl": "https://github.com/timo-cavelius/StylisticBias", "url": "https://arxiv.org/abs/2606.20527"}], "scope": "benchmark_repo", "url": "https://github.com/timo-cavelius/StylisticBias", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_stylisticbias_ee73496d", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/timo-cavelius/StylisticBias", "data": null, "hfPaper": "https://huggingface.co/papers/2606.20527", "paper": "https://arxiv.org/abs/2606.20527", "pdf": "https://arxiv.org/pdf/2606.20527", "project": "https://hf.co/datasets/shaghayegh/stylistic-bias-dataset", "report": "https://arxiv.org/abs/2606.20527"}, "metrics": [], "motivation": "Multimodal large language models (MLLMs) are increasingly deployed in personally and societally consequential settings, yet the visual cues that shape how these models judge people remain poorly understood.", "name": "StylisticBias", "oneLine": "StylisticBias is a benchmark of 25,000 images with controlled single-attribute variations to evaluate attribute-level social bias in multimodal LLMs. It measures how visual cues shift model judgments in 25 scenarios.", "paperTitle": "StylisticBias: A Few Human Visual Cues Drive Most Social Biases in MLLMs", "primaryDomain": "General AI", "publication": {"evidence": "Accepted to the non-archival workshops AI4Good and Culture x AI at ICML 2026", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2606.20527", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "non-archival workshops AI4Good and Culture x AI at ICML 2026", "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "StylisticBias Team", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/timo-cavelius/StylisticBias"}], "ranking": {}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-18", "source": {"authors": ["Shaghayegh Kolli", "Timo Cavelius", "Nafiseh Nikeghbal", "Samantha Dalal", "Jana Diesner"], "categories": ["cs.CL", "cs.CV"], "id": "2606.20527", "title": "StylisticBias: A Few Human Visual Cues Drive Most Social Biases in MLLMs", "type": "arxiv", "url": "https://arxiv.org/abs/2606.20527"}, "sourceUpdatedAt": "2026-06-18T17:39:56Z", "topics": ["cs.CL"], "venueAttempts": [{"decisionRaw": "Accepted to the non-archival workshops AI4Good and Culture x AI at ICML 2026", "evidence": [{"level": "author-claim", "observedAt": "2026-10-14T01:46:58.058199Z", "rawValue": "Accepted to the non-archival workshops AI4Good and Culture x AI at ICML 2026", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2606.20527"}], "reviewStatus": "accepted", "venueName": "non-archival workshops AI4Good and Culture x AI at ICML 2026"}], "whyItMatters": "Social bias in multimodal models is often confounded by identity. StylisticBias isolates visual cues, enabling targeted bias diagnosis and mitigation."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
