{"record": {"aliases": ["Knowing What Not to Answer: Selective Non-Compliance in Vision-Language Models"], "artifacts": [{"id": "arxiv:2609.04720", "kind": "paper", "url": "https://arxiv.org/abs/2609.04720"}, {"id": "gh:mz-kim/kona", "kind": "repo", "url": "https://github.com/mz-kim/KoNA"}], "authors": ["Minji Kim", "Jihyoung Jang", "Hyounghun Kim"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.04720", "source_url": "https://arxiv.org/abs/2609.04720"}, "categories": ["General AI", "Multimodal", "Safety", "Safety & Trustworthiness"], "description": {"en": "KoNA evaluates selective non-compliance in vision-language models across five categories: False Premise, Visual Inaccessibility, Universal Unknown, Task Feasibility, and Safety. Each task measures query-level and component-level non-compliance using paired single and compound queries."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.04720", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.04720", "title": "KoNA"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2609.04720", "modality": null, "name": "KoNA", "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_knowing-what-not-to-answer-selective-non-c_abb19683", "origin_source": "arxiv", "recognition_confidence": "0.7", "record_sha256": "ec7018d6c394a40567a19e8e719bab124758ab3f85d995f4691adcadca2fb7ca", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-08T00:56:29.478814Z", "source_url": "https://arxiv.org/abs/2609.04720"}, "publisher": null, "released": "2026-09-04", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.04720", "source_url": "https://arxiv.org/abs/2609.04720"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-04720", "source": "claire_radar", "source_benchmark_id": "2609.04720", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Knowing What Not to Answer: Selective Non-Compliance in Vision-Language Models"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-09-07", "attemptedAt": "2026-09-07T01:28:43.988909Z", "githubRepo": "https://github.com/mz-kim/KoNA", "githubScope": "benchmark_repo", "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2609.04720", "observedAt": "2026-09-06T01:30:14.942019Z", "signalStatus": {"githubStars": {"checkedAt": "2026-09-07T01:28:43.988909Z", "state": "unavailable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-09-07T01:28:43.988909Z", "state": "unavailable"}}, "status": "unavailable"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The benchmark addresses a novel and practical aspect of VLM safety with a released dataset and code, targeting a broad audience in multimodal AI research.", "score": 75}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-08T08:59:06.803818Z", "inputHash": "7a08b1229bbbe6540719bd829d76d0d5df587a2d4943e276a6daf1f08e3ee4ea", "model": "deepseek-v4-flash"}, "curation": {"canonicalNameEvidence": "we introduce KoNA, a benchmark for evaluating selective non-compliance in VLMs", "canonicalNameSource": "abstract", "decisionReason": "The benchmark is released with code, datasets on Hugging Face, and evaluation scripts, enabling public reuse and comparison.", "model": "deepseek-v4-pro", "reviewedAt": "2026-09-08T00:56:29.478814Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "KoNA evaluates selective non-compliance in vision-language models across five categories: False Premise, Visual Inaccessibility, Universal Unknown, Task Feasibility, and Safety. Each task measures query-level and component-level non-compliance using paired single and compound queries.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "In this paper, we introduce KoNA, a benchmark for evaluating selective non-compliance in VLMs across five categories: False Premise, Visual Inaccessibility, Universal Unknown, Task Feasibility, and Safety."}, "familyId": "bmf_1f06e3511e03", "firstSeenAt": "2026-09-07", "heat": null, "id": "bm_knowing-what-not-to-answer-selective-non-c_abb19683", "indexedAt": "2026-09-07T00:39:11.978130Z", "industrySectors": [], "links": {"code": "https://github.com/mz-kim/KoNA", "data": null, "paper": "https://arxiv.org/abs/2609.04720", "pdf": "https://arxiv.org/pdf/2609.04720", "project": null, "report": "https://arxiv.org/abs/2609.04720"}, "metrics": [], "motivation": "Vision-language models (VLMs) are expected to respond helpfully to appropriate requests while withholding compliance with requests that are incorrect, unsafe, infeasible, or unanswerable.", "name": "KoNA", "oneLine": "KoNA evaluates selective non-compliance in vision-language models across five categories: False Premise, Visual Inaccessibility, Universal Unknown, Task Feasibility, and Safety. Each task measures query-level and component-level non-compliance using paired single and compound queries.", "paperTitle": "Knowing What Not to Answer: Selective Non-Compliance in Vision-Language Models", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.04720", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-07T00:39:11.978130Z"}, "publications": [], "publishers": [{"name": "mz-kim", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/mz-kim/KoNA"}], "ranking": {"30d": {"confidence": "Low", "coverage": 0.0, "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": null, "value": null}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v13", "rank": null, "score": null}, "90d": {"confidence": "Low", "coverage": 0.0, "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": null, "value": null}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v13", "rank": null, "score": null}}, "readiness": "Runnable", "recognitionConfidence": 0.7, "relation": "introduces", "releasedAt": "2026-09-04", "source": {"authors": ["Minji Kim", "Jihyoung Jang", "Hyounghun Kim"], "categories": ["cs.CL", "cs.AI"], "id": "2609.04720", "title": "Knowing What Not to Answer: Selective Non-Compliance in Vision-Language Models", "type": "arxiv", "url": "https://arxiv.org/abs/2609.04720"}, "sourceUpdatedAt": "2026-09-04T04:45:26Z", "topics": ["Multimodal", "Safety"], "venueAttempts": [], "whyItMatters": "Existing benchmarks treat queries as wholly compliant or non-compliant, missing real-world mixed queries. KoNA provides a reusable testbed to assess whether VLMs can refuse only inappropriate components while remaining helpful, informing safety and reliability improvements in multimodal assistants."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
