{"record": {"aliases": ["ContextBias: Controlled Evaluation of Bias Persistence Under Context Shift in Text-to-Image Models"], "artifacts": [{"id": "arxiv:2608.29847", "kind": "paper", "url": "https://arxiv.org/abs/2608.29847"}, {"id": "gh:sina-emami/contextbias", "kind": "repo", "url": "https://github.com/Sina-Emami/ContextBias"}, {"id": "hf:shaghayegh/contextbias", "kind": "dataset", "url": "https://huggingface.co/datasets/shaghayegh/ContextBias"}], "categories": ["General AI", "Safety & Trustworthiness", "cs.CV"], "description": {"en": "Evaluates text-to-image models on 92 roles and 1,656 controlled prompts across context-free, related, and unrelated conditions, measuring bias persistence in visual attributes through metrics like Bias Intensity and Context Consistency Score."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.29847", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.29847", "title": "ContextBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2608.29847", "modality": null, "name": "ContextBench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_contextbias-controlled-evaluation-of-bias-_978cb4aa", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "b205feb6986fe68146acbf749f376b3434be7109aa274a5f4b923d9c71a1910d", "relation": "evaluates_only", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-02T05:23:11.715633Z", "source_url": "https://arxiv.org/abs/2608.29847"}, "publisher": null, "released": "2026-08-30", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.29847", "source_url": "https://arxiv.org/abs/2608.29847"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-29847", "source": "claire_radar", "source_benchmark_id": "2608.29847", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["ContextBias: Controlled Evaluation of Bias Persistence Under Context Shift in Text-to-Image Models"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-09-02", "attemptedAt": "2026-09-02T05:12:51.621571Z", "githubRepo": "https://github.com/Sina-Emami/ContextBias", "githubScope": "benchmark_repo", "githubStars": 0, "hfDailySubmittedAt": "2026-09-01T00:00:00.000Z", "hfDataset": "shaghayegh/ContextBias", "hfDatasetDownloads": 496, "hfDatasetLikes": 0, "hfPaperUpvotes": 4, "hfPaperUrl": "https://huggingface.co/papers/2608.29847", "observedAt": "2026-09-02T05:12:51.621571Z", "signalStatus": {"githubStars": {"observedAt": "2026-09-02T05:12:51.621571Z", "state": "fresh"}, "hfDatasetDownloads": {"observedAt": "2026-09-02T05:12:51.621571Z", "state": "fresh"}, "hfDatasetLikes": {"observedAt": "2026-09-02T05:12:51.621571Z", "state": "fresh"}, "hfPaperUpvotes": {"observedAt": "2026-09-02T05:12:51.621571Z", "state": "fresh"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The benchmark addresses a timely bias evaluation gap in text-to-image models and provides full public code and a dataset, which typically draws immediate interest from fairness-focused researchers.", "score": 67}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-02T05:23:11.715633Z", "inputHash": "9f081e8b13997ed755e010d660f80fa27305c78576b213b7a6de93fca980ec6b", "model": "deepseek-v4-pro"}, "curation": {"canonicalNameEvidence": "We introduce ContextBias, a controlled evaluation framework, and ContextBench, a benchmark spanning 92 roles and 1,656 semantically controlled prompts, designed to isolate the effect of contextual variation on role-linked visual representations.", "canonicalNameSource": "abstract", "decisionReason": "Formally named benchmark with 1,656 prompts, clear metrics, and public code and dataset at provided repositories.", "model": "deepseek-v4-pro", "reviewedAt": "2026-09-02T05:23:11.715633Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates text-to-image models on 92 roles and 1,656 controlled prompts across context-free, related, and unrelated conditions, measuring bias persistence in visual attributes through metrics like Bias Intensity and Context Consistency Score.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "public artifact URL", "named entity is a model, method, framework, or dataset used in evaluation"], "snippet": "We introduce ContextBias, a controlled evaluation framework, and ContextBench, a benchmark spanning 92 roles and 1,656 semantically controlled prompts, designed to isolate the effect of contextual variation on role-linked visual representations."}, "familyId": "bmf_2af32df0f797", "firstSeenAt": "2026-09-01", "heat": null, "id": "bm_contextbias-controlled-evaluation-of-bias-_978cb4aa", "indexedAt": "2026-09-01T06:28:29.895112Z", "industrySectors": [], "links": {"code": "https://github.com/Sina-Emami/ContextBias", "data": "https://huggingface.co/datasets/shaghayegh/ContextBias", "paper": "https://arxiv.org/abs/2608.29847", "pdf": "https://arxiv.org/pdf/2608.29847", "project": null, "report": "https://arxiv.org/abs/2608.29847"}, "metrics": [], "motivation": "Text-to-image models learn associations between concepts - in the case of this paper, people's professions, which we refer to as roles - and visual attributes.", "name": "ContextBench", "oneLine": "Evaluates text-to-image models on 92 roles and 1,656 controlled prompts across context-free, related, and unrelated conditions, measuring bias persistence in visual attributes through metrics like Bias Intensity and Context Consistency Score.", "paperTitle": "ContextBias: Controlled Evaluation of Bias Persistence Under Context Shift in Text-to-Image Models", "primaryDomain": "General AI", "publication": {"evidence": "Accepted to EMNLP 2026", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2608.29847", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "EMNLP 2026", "verifiedAt": "2026-09-01T06:28:29.895112Z"}, "publications": [], "publishers": [{"name": "Technical University of Munich", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/Sina-Emami/ContextBias"}], "ranking": {"30d": {"confidence": "High", "coverage": 1.0, "datasetDownloadRank": 9, "datasetRankPopulation": 34, "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": 30}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.0, "value": 0}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": 0.6479868386415211, "value": 496}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.285264874482679, "value": 4}}, "confidence": "High", "coverage": 1.0, "rank": 186, "score": 18}, "method": "attention-ranking-v13", "rank": 186, "score": 18}, "90d": {"confidence": "High", "coverage": 1.0, "datasetDownloadRank": 22, "datasetRankPopulation": 70, "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": 0.0, "value": 0}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": 0.5142739076015804, "value": 496}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.285264874482679, "value": 4}}, "confidence": "High", "coverage": 1.0, "rank": 403, "score": 20}, "method": "attention-ranking-v13", "rank": 403, "score": 20}}, "readiness": "Runnable", "recognitionConfidence": 0.55, "relation": "evaluates_only", "releasedAt": "2026-08-30", "source": {"authors": ["Shaghayegh Kolli", "Sina Emami", "Moreno D'Inc\\`a", "Pouyan Nejadi", "Nicu Sebe", "Massimiliano Mancini", "Jana Diesner"], "categories": ["cs.CV", "cs.CL"], "id": "2608.29847", "title": "ContextBias: Controlled Evaluation of Bias Persistence Under Context Shift in Text-to-Image Models", "type": "arxiv", "url": "https://arxiv.org/abs/2608.29847"}, "sourceUpdatedAt": "2026-08-30T15:23:07Z", "topics": ["cs.CV"], "venueAttempts": [{"decisionRaw": "Accepted to EMNLP 2026", "evidence": [{"level": "author-claim", "observedAt": "2026-09-01T06:28:29.895112Z", "rawValue": "Accepted to EMNLP 2026", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2608.29847"}], "reviewStatus": "accepted", "venueName": "EMNLP 2026"}], "whyItMatters": "Provides a controlled way to detect stereotype persistence that context-free bias benchmarks miss, with a public dataset and code that lets other teams reproduce and compare evaluations."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
