{"record": {"aliases": ["Learning to Deny: Action Denial in Multimodal Large Language Models"], "artifacts": [{"id": "arxiv:2606.31187", "kind": "paper", "url": "https://arxiv.org/abs/2606.31187"}, {"id": "gh:raiyaan-abdullah/learn-to-deny", "kind": "repo", "url": "https://github.com/raiyaan-abdullah/Learn-to-Deny"}], "authors": ["Raiyaan Abdullah", "Shehreen Azad", "Yogesh Singh Rawat"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.31187", "source_url": "https://arxiv.org/abs/2606.31187"}, "categories": ["General AI", "Multimodal"], "description": {"en": "A paired-video benchmark evaluating whether multimodal language models can distinguish action presence from action denial under matched contextual and motion cues."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.31187", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.31187", "title": "UCF101-AD"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2606.31187", "modality": null, "name": "UCF101-AD", "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_learning-to-deny-action-denial-in-multimod_ec508261", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "df7ec8db3b30a678cd0e7d92d9b765703a8c9cd6b6b3b08554380736eafca0e4", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T20:25:40.535027Z", "source_url": "https://arxiv.org/abs/2606.31187"}, "publisher": null, "released": "2026-06-30", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.31187", "source_url": "https://arxiv.org/abs/2606.31187"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-31187", "source": "claire_radar", "source_benchmark_id": "2606.31187", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Learning to Deny: Action Denial in Multimodal Large Language Models"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Multimodal", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The benchmark addresses a clear failure mode in multimodal video reasoning and evaluates 20 contemporary models.", "score": 4}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T20:25:40.535027Z", "inputHash": "670d82bc20f2ee63aebf040091ba34ab61c1c80d060fecf6d9a93ad08c00d3f4", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We introduce UCF101-AD, a large-scale benchmark consisting of paired Action-Presence and Action-Denial clips", "canonicalNameSource": "abstract", "decisionReason": "The source formally names a reusable benchmark, defines its denial-evaluation task, reports accuracy as a concrete metric, and supplies a public dataset-and-code link.", "evaluationEvidence": {"reuse": "Dataset and relevant code: https://github.com/raiyaan-abdullah/Learn-to-Deny.", "scoring": "models that exceed 85% accuracy on the positive action classes collapse below 50% on their action-denial counterparts", "task": "designed to evaluate this capacity for denial"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T20:25:40.535027Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A paired-video benchmark evaluating whether multimodal language models can distinguish action presence from action denial under matched contextual and motion cues.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "public artifact URL"], "snippet": "We introduce UCF101-AD, a large-scale benchmark consisting of paired Action-Presence and Action-Denial clips, designed to evaluate this capacity for denial."}, "familyId": "bmf_041f887fa861", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_learning-to-deny-action-denial-in-multimod_ec508261", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/raiyaan-abdullah/Learn-to-Deny", "data": null, "paper": "https://arxiv.org/abs/2606.31187", "pdf": "https://arxiv.org/pdf/2606.31187", "project": null, "report": "https://arxiv.org/abs/2606.31187"}, "metrics": [], "motivation": "Multimodal large language models (MLLMs) have rapidly advanced video understanding, achieving strong zero-shot and few-shot recognition across standard benchmarks.", "name": "UCF101-AD", "oneLine": "A paired-video benchmark evaluating whether multimodal language models can distinguish action presence from action denial under matched contextual and motion cues.", "paperTitle": "Learning to Deny: Action Denial in Multimodal Large Language Models", "primaryDomain": "General AI", "publication": {"evidence": "Accepted to ECCV 2026 main conference", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2606.31187", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "ECCV 2026 main conference", "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "publishers": [{"name": "raiyaan-abdullah", "organizationType": "community", "role": "benchmark-publisher", "sourceUrl": "https://github.com/raiyaan-abdullah/Learn-to-Deny"}], "readiness": "Runnable", "recognitionConfidence": 0.55, "relation": "introduces", "releasedAt": "2026-06-30", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Raiyaan Abdullah", "Shehreen Azad", "Yogesh Singh Rawat"], "categories": ["cs.CV"], "id": "2606.31187", "title": "Learning to Deny: Action Denial in Multimodal Large Language Models", "type": "arxiv", "url": "https://arxiv.org/abs/2606.31187"}, "sourceUpdatedAt": "2026-06-30T06:16:17Z", "topics": ["Multimodal"], "venueAttempts": [{"decisionRaw": "Accepted to ECCV 2026 main conference", "evidence": [{"level": "author-claim", "observedAt": "2026-08-19T11:05:13.395391Z", "rawValue": "Accepted to ECCV 2026 main conference", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2606.31187"}], "reviewStatus": "accepted", "venueName": "ECCV 2026 main conference"}], "whyItMatters": "It measures whether video models verify that an action occurs instead of affirming an action suggested by its surrounding context."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
