{"record": {"aliases": ["Large Language Models Systematically Favor Popular Options: Evidence and Mitigation Across MCQs"], "artifacts": [{"id": "arxiv:2608.29257", "kind": "paper", "url": "https://arxiv.org/abs/2608.29257"}, {"id": "gh:datascienceuibk/popmcq", "kind": "repo", "url": "https://github.com/DataScienceUIBK/PopMCQ"}], "authors": ["Abdelrahman Abdallah", "Mohammed Ali", "Bhawna Piryani", "Mahmoud Abdalla", "Adam Jatowt"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.29257", "source_url": "https://arxiv.org/abs/2608.29257"}, "categories": ["Cybersecurity", "Language & Knowledge", "cs.CL"], "description": {"en": "Evaluates multiple-choice question answering across 73,260 MCQ items with six strategies that systematically vary distractor popularity, using accuracy, HPSR, and other bias metrics."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.29257", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.29257", "title": "PopMCQ"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2608.29257", "modality": null, "name": "PopMCQ", "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_large-language-models-systematically-favor_639da924", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "ad42379f462a8928e8f52be7f876e336de12c421333c3b8cc39e983a95a1bc54", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-02T05:23:11.715633Z", "source_url": "https://arxiv.org/abs/2608.29257"}, "publisher": null, "released": "2026-08-29", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.29257", "source_url": "https://arxiv.org/abs/2608.29257"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-29257", "source": "claire_radar", "source_benchmark_id": "2608.29257", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Large Language Models Systematically Favor Popular Options: Evidence and Mitigation Across MCQs"], "annotation": "Unknown", "applicationDomains": ["Cybersecurity"], "area": "Language & Knowledge", "attention": {"asOf": "2026-09-02", "attemptedAt": "2026-09-02T05:12:51.621571Z", "githubRepo": "https://github.com/DataScienceUIBK/PopMCQ", "githubScope": "benchmark_repo", "githubStars": 1, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2608.29257", "observedAt": "2026-09-02T05:12:51.621571Z", "signalStatus": {"githubStars": {"observedAt": "2026-09-02T05:12:51.621571Z", "state": "fresh"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-09-02T05:12:51.621571Z", "state": "unavailable"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The benchmark targets a common evaluation format and includes a debiasing method, with a well-documented public repository that is likely to generate immediate interest in the NLP community.", "score": 78}, "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": "fae9f843bbd7f5cc0db4536f6931c4aea8dd53fd289cea7defd2162298397c98", "model": "deepseek-v4-pro"}, "curation": {"canonicalNameEvidence": "To systematically isolate this phenomenon, we introduce \\textbf{PopMCQ}, a benchmark with six controlled strategies that vary option popularity while keeping the correct answer fixed.", "canonicalNameSource": "abstract", "decisionReason": "Formally named benchmark with controlled strategies, public dataset on Hugging Face, and code that supports evaluation of new models.", "model": "deepseek-v4-pro", "reviewedAt": "2026-09-02T05:23:11.715633Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates multiple-choice question answering across 73,260 MCQ items with six strategies that systematically vary distractor popularity, using accuracy, HPSR, and other bias metrics.", "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"], "snippet": "To systematically isolate this phenomenon, we introduce \\textbf{PopMCQ}, a benchmark with six controlled strategies that vary option popularity while keeping the correct answer fixed."}, "familyId": "bmf_f458f72b82eb", "firstSeenAt": "2026-09-01", "heat": null, "id": "bm_large-language-models-systematically-favor_639da924", "indexedAt": "2026-09-01T06:28:29.895112Z", "industrySectors": ["Cybersecurity"], "links": {"code": "https://github.com/DataScienceUIBK/PopMCQ", "data": null, "paper": "https://arxiv.org/abs/2608.29257", "pdf": "https://arxiv.org/pdf/2608.29257", "project": null, "report": "https://arxiv.org/abs/2608.29257"}, "metrics": [], "motivation": "Multiple-choice questions (MCQs) are a standard format for evaluating large language models (LLMs), yet the popularity of answer options can confound evaluation.", "name": "PopMCQ", "oneLine": "Evaluates multiple-choice question answering across 73,260 MCQ items with six strategies that systematically vary distractor popularity, using accuracy, HPSR, and other bias metrics.", "paperTitle": "Large Language Models Systematically Favor Popular Options: Evidence and Mitigation Across MCQs", "primaryDomain": "Cybersecurity", "publication": {"evidence": "Accepted at MAIN EMNLP 2026", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2608.29257", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "MAIN EMNLP 2026", "verifiedAt": "2026-09-01T06:28:29.895112Z"}, "publications": [], "publishers": [{"name": "Data Science UIBK", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/DataScienceUIBK/PopMCQ"}], "ranking": {"30d": {"confidence": "Low", "coverage": 0.55, "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": 0.10323645750017778, "value": 1}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.55, "rank": 94, "score": 28}, "method": "attention-ranking-v13", "rank": 94, "score": 28}, "90d": {"confidence": "Low", "coverage": 0.55, "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": 0.09434039834931938, "value": 1}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.55, "rank": 257, "score": 28}, "method": "attention-ranking-v13", "rank": 257, "score": 28}}, "readiness": "Runnable", "recognitionConfidence": 0.55, "relation": "introduces", "releasedAt": "2026-08-29", "source": {"authors": ["Abdelrahman Abdallah", "Mohammed Ali", "Bhawna Piryani", "Mahmoud Abdalla", "Adam Jatowt"], "categories": ["cs.CL"], "id": "2608.29257", "title": "Large Language Models Systematically Favor Popular Options: Evidence and Mitigation Across MCQs", "type": "arxiv", "url": "https://arxiv.org/abs/2608.29257"}, "sourceUpdatedAt": "2026-08-29T13:21:07Z", "topics": ["cs.CL"], "venueAttempts": [{"decisionRaw": "Accepted at MAIN EMNLP 2026", "evidence": [{"level": "author-claim", "observedAt": "2026-09-01T06:28:29.895112Z", "rawValue": "Accepted at MAIN EMNLP 2026", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2608.29257"}], "reviewStatus": "accepted", "venueName": "MAIN EMNLP 2026"}], "whyItMatters": "Exposes a systematic popularity bias in MCQ evaluation and offers a reusable dataset and inference-time debiasing method, enabling fairer model comparisons."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
