{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.05874", "kind": "paper", "url": "https://arxiv.org/abs/2606.05874"}], "authors": ["Huiyuan Zheng", "Houtao Zhang", "Boyang Wang", "Qingyi Si", "Hongcheng Guo"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.05874", "source_url": "https://arxiv.org/abs/2606.05874"}, "categories": ["General AI", "Multimodal", "Safety & Trustworthiness"], "description": {"en": "RandomBench evaluates whether multimodal LLMs maintain distributionally neutral behavior when selecting among equivalent options, providing metrics for entropy and distributional bias under explicit random instructions."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.05874", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.05874", "title": "RandomBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2606.05874", "modality": null, "name": "RandomBench", "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_randombench_dfec6d0b", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "2d1bdde5edce7251d283c8ed8eb0db776573de0573cd7a33b56fb05d02815a56", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.05874"}, "publisher": null, "released": "2026-06-04", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.05874", "source_url": "https://arxiv.org/abs/2606.05874"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-05874", "source": "claire_radar", "source_benchmark_id": "2606.05874", "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/2606.05874", "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": [], "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": "deb5a31edd9729ff7bcf690a98368458e859d96d602081a530e7924359a79475", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "RandomBench evaluates whether multimodal LLMs maintain distributionally neutral behavior when selecting among equivalent options, providing metrics for entropy and distributional bias under explicit random instructions.", "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 bridge this gap, we propose RandomBench, a benchmark designed to evaluate whether MLLMs can maintain distributionally neutral behavior when selecting among equivalent options."}, "familyId": "bmf_6cb8c3fc76bb", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_randombench_dfec6d0b", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2606.05874", "paper": "https://arxiv.org/abs/2606.05874", "pdf": "https://arxiv.org/pdf/2606.05874", "project": null, "report": "https://arxiv.org/abs/2606.05874"}, "metrics": [], "motivation": "Current evaluations for Multimodal Large Language Models (MLLMs) overwhelmingly focus on utility-driven objectives, leaving model behavior under logic-neutral scenarios largely underexplored.", "name": "RandomBench", "oneLine": "RandomBench evaluates whether multimodal LLMs maintain distributionally neutral behavior when selecting among equivalent options, providing metrics for entropy and distributional bias under explicit random instructions.", "paperTitle": "Evaluating Stochastic Collapse and Implicit Bias in Multimodal Large Language Models", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.05874", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-04", "source": {"authors": ["Huiyuan Zheng", "Houtao Zhang", "Boyang Wang", "Qingyi Si", "Hongcheng Guo"], "categories": ["cs.CL"], "id": "2606.05874", "title": "Evaluating Stochastic Collapse and Implicit Bias in Multimodal Large Language Models", "type": "arxiv", "url": "https://arxiv.org/abs/2606.05874"}, "sourceUpdatedAt": "2026-06-04T08:47:15Z", "topics": ["Multimodal"], "venueAttempts": [], "whyItMatters": "Logic-neutral scenarios are underexplored in MLLM evaluation. RandomBench introduces a way to quantify stochastic collapse, a bias toward non-uniform choices that affects repetitive behavior and coverage, aiding design of more robust models."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
