{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.11063", "kind": "paper", "url": "https://arxiv.org/abs/2606.11063"}], "authors": ["Joachim Schaeffer", "Alexander Panfilov", "Thomas Jiralerspong", "Guillaume Lajoie", "Jonas Geiping", "Yoshua Bengio", "Roland S. Zimmermann"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.11063", "source_url": "https://arxiv.org/abs/2606.11063"}, "categories": ["General AI", "Language & Knowledge", "cs.AI"], "description": {"en": "CIAware-Bench measures control intervention awareness in language models through four task domains: essay writing, BigCodeBench, Bash Arena, and SHADE-Arena. Models are tested on distinguishing their own trajectories from those modified by a control protocol, with variations in watermarking, side-task presence, and protocol."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.11063", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.11063", "title": "CIAware-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Bernstein Center for Computational Neuroscience Tübingen", "ELLIS Institute Tübingen", "Mila - Quebec Artificial Intelligence Institute", "Université de Montréal", "Google DeepMind (United Kingdom)"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.11063", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.11063", "source_url": "https://openalex.org/W7164171296"}, "key": "claire-radar:2606.11063", "modality": null, "name": "CIAware-Bench", "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_ciaware-bench_e29a610f", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "b0916b75fa9151acce90b8973f4153bfd80723b58f5087e6719e78326ba69784", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.11063"}, "publisher": null, "released": "2026-06-09", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.11063", "source_url": "https://arxiv.org/abs/2606.11063"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-11063", "source": "claire_radar", "source_benchmark_id": "2606.11063", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "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/2606.11063", "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": "69e83acb4b5b3c483e323308f788839786750feff0a7b9062cfa8af73f86a2db", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "CIAware-Bench measures control intervention awareness in language models through four task domains: essay writing, BigCodeBench, Bash Arena, and SHADE-Arena. Models are tested on distinguishing their own trajectories from those modified by a control protocol, with variations in watermarking, side-task presence, and protocol.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We introduce \\textbf{CIAware-Bench}, a benchmark for measuring \\textbf{c}ontrol \\textbf{i}ntervention (CI) awareness across frontier models."}, "familyId": "bmf_78a7e9aa9db3", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_ciaware-bench_e29a610f", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2606.11063", "paper": "https://arxiv.org/abs/2606.11063", "pdf": "https://arxiv.org/pdf/2606.11063", "project": null, "report": "https://arxiv.org/abs/2606.11063"}, "metrics": [], "motivation": "AI control protocols oversee untrusted models by monitoring their actions and modifying potentially unsafe steps, often using a trusted model.", "name": "CIAware-Bench", "oneLine": "CIAware-Bench measures control intervention awareness in language models through four task domains: essay writing, BigCodeBench, Bash Arena, and SHADE-Arena. Models are tested on distinguishing their own trajectories from those modified by a control protocol, with variations in watermarking, side-task presence, and protocol.", "paperTitle": "CIAware-Bench: Benchmarking Control Intervention Awareness Across Frontier LLMs", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.11063", "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": "2026-06-09", "source": {"authors": ["Joachim Schaeffer", "Thomas Jiralerspong", "Alexander Panfilov", "Guillaume Lajoie", "Jonas Geiping", "Yoshua Bengio", "Roland S. Zimmermann"], "categories": ["cs.AI", "cs.LG"], "id": "2606.11063", "title": "CIAware-Bench: Benchmarking Control Intervention Awareness Across Frontier LLMs", "type": "arxiv", "url": "https://arxiv.org/abs/2606.11063"}, "sourceUpdatedAt": "2026-06-09T16:24:16Z", "topics": ["cs.AI"], "venueAttempts": [], "whyItMatters": "This benchmark addresses the evaluation gap of determining whether models can detect modifications to their trajectories, which is critical for AI control protocols. It provides practical value by informing the design of interventions that are harder for models to detect, thereby improving the robustness of AI oversight systems."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
