{"record": {"aliases": ["Unified Hallucination Fuzzing for Multimodal Large Language Models"], "artifacts": [{"id": "arxiv:2608.07525", "kind": "paper", "url": "https://arxiv.org/abs/2608.07525"}, {"id": "gh:lancezpf/evalhall", "kind": "repo", "url": "https://github.com/LanceZPF/EvalHall"}], "authors": ["Pengfei Zhou", "Jiajun Song", "Zhiwei Tang", "Yixing Ma", "Xiaopeng Peng", "Donghui Si", "Yuhang Xu", "Huiqi Song", "Yiyuan Miao", "Yichen Qian", "Weihua Chen", "Wangbo Zhao", "Bohan Zhuang", "Jiasheng Tang", "Yang You"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.07525", "source_url": "https://arxiv.org/abs/2608.07525"}, "categories": ["General AI", "Multimodal"], "description": {"en": "UniHall evaluates multimodal-model hallucinations across object, instruction, and knowledge dimensions, alongside adaptive fuzzing stress tests."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.07525", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.07525", "title": "UniHall"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2608.07525", "modality": null, "name": "UniHall", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-09-25T04:40:22Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_unified-hallucination-fuzzing-for-multimod_1787cf6e", "origin_source": "arxiv", "recognition_confidence": "0.7", "record_sha256": "4302458d45b2d960bf24dc911a3e7ca37a478205f2b00a80310b2d8f3dc4ae90", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T17:28:19.823139Z", "source_url": "https://arxiv.org/abs/2608.07525"}, "publisher": null, "released": "2026-07-15", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.07525", "source_url": "https://arxiv.org/abs/2608.07525"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-07525", "source": "claire_radar", "source_benchmark_id": "2608.07525", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Unified Hallucination Fuzzing for Multimodal Large Language Models"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Multimodal", "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "Multimodal hallucination and adaptive stress testing are active topics, and the public implementation supports near-term reuse.", "score": 62}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T17:28:19.823139Z", "inputHash": "4049a748b0367671a476e228d23378de29cc5ca87e3d9a9ee9c6446bdec8a055", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "we introduce UniHall, a fine-grained dataset grounded in a unified taxonomy", "canonicalNameSource": "abstract", "decisionReason": "The supplied abstract formally names the evaluation, states its task taxonomy and structured metric suite, and provides a public repository for the framework, code, and benchmark.", "evaluationEvidence": {"reuse": "The framework, code and benchmark are available at https://github.com/LanceZPF/EvalHall.", "scoring": "SAMF incorporates a structured metric suite driven by an ensemble of multi-modal oracles", "task": "a unified taxonomy spanning Object, Instruction, and Knowledge dimensions"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T17:28:19.823139Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "UniHall evaluates multimodal-model hallucinations across object, instruction, and knowledge dimensions, alongside adaptive fuzzing stress tests.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "To bridge this gap, we present a systematic evaluation framework integrating a comprehensive benchmark with self-evolving stress testing."}, "familyId": "bmf_749ec6eb601a", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_unified-hallucination-fuzzing-for-multimod_1787cf6e", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/LanceZPF/EvalHall", "data": null, "paper": "https://arxiv.org/abs/2608.07525", "pdf": "https://arxiv.org/pdf/2608.07525", "project": null, "report": "https://arxiv.org/abs/2608.07525"}, "metrics": [], "motivation": "Hallucination remains a persistent challenge for Multimodal Large Language Models (MLLMs), severely limiting their reliability in high-stakes applications.", "name": "UniHall", "oneLine": "UniHall evaluates multimodal-model hallucinations across object, instruction, and knowledge dimensions, alongside adaptive fuzzing stress tests.", "paperTitle": "Unified Hallucination Fuzzing for Multimodal Large Language Models", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.07525", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Runnable", "recognitionConfidence": 0.7, "relation": "introduces", "releasedAt": "2026-07-15", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Pengfei Zhou", "Jiajun Song", "Zhiwei Tang", "Yixing Ma", "Xiaopeng Peng", "Donghui Si", "Yuhang Xu", "Huiqi Song", "Yiyuan Miao", "Yichen Qian", "Weihua Chen", "Wangbo Zhao", "Bohan Zhuang", "Jiasheng Tang", "Yang You"], "categories": ["cs.CL", "cs.AI"], "id": "2608.07525", "title": "Unified Hallucination Fuzzing for Multimodal Large Language Models", "type": "arxiv", "url": "https://arxiv.org/abs/2608.07525"}, "sourceUpdatedAt": "2026-07-15T17:57:37Z", "topics": ["Multimodal"], "venueAttempts": [], "whyItMatters": "It combines conventional benchmark evaluation with evolving adversarial inputs intended to expose robustness failures hidden by saturated static tests."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
