{"record": {"aliases": ["SchemeArena: Factorized Stress Testing of Scheming in LLM Agents"], "artifacts": [{"id": "arxiv:2609.08126", "kind": "paper", "url": "https://arxiv.org/abs/2609.08126"}, {"id": "gh:launchnlp/schemearena", "kind": "repo", "url": "https://github.com/launchnlp/SchemeArena"}], "authors": ["Jie Ruan", "Inderjeet Nair", "Amy Liu", "Muhammad Khalifa", "Yusheng Zhou", "Lu Wang"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.08126", "source_url": "https://arxiv.org/abs/2609.08126"}, "categories": ["General AI", "Safety", "Safety & Trustworthiness"], "description": {"en": "Evaluates scheming in LLM agents across 400 factorized scenarios, with a monitor that scores covert misaligned behavior along five rubric dimensions."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.08126", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.08126", "title": "SchemeArena"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["University of Michigan"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2609.08126", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2609.08126", "source_url": "https://openalex.org/W7212020693"}, "key": "claire-radar:2609.08126", "modality": null, "name": "SchemeArena", "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": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_schemearena-factorized-stress-testing-of-s_9000c40f", "origin_source": "arxiv", "recognition_confidence": "0.7", "record_sha256": "b2c833290257140e3a84d7c62cdb76fab53b1f87e6ca3c846d07174d1f9dd5a0", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-10T00:52:28.558291Z", "source_url": "https://arxiv.org/abs/2609.08126"}, "publisher": null, "released": "2026-09-08", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.08126", "source_url": "https://arxiv.org/abs/2609.08126"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-08126", "source": "claire_radar", "source_benchmark_id": "2609.08126", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["SchemeArena: Factorized Stress Testing of Scheming in LLM Agents"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-09-09", "attemptedAt": "2026-09-09T06:48:27.347268Z", "githubRepo": "https://github.com/launchnlp/SchemeArena", "githubScope": "benchmark_repo", "githubStars": 0, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2609.08126", "observedAt": "2026-09-09T06:48:27.347268Z", "signalStatus": {"githubStars": {"observedAt": "2026-09-09T06:48:27.347268Z", "state": "fresh"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-09-09T06:48:27.347268Z", "state": "unavailable"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The benchmark tackles a high-stakes safety topic with a factorized scenario framework and open-source tooling, which may draw attention from AI safety and agent evaluation communities.", "score": 60}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-10T00:52:28.558291Z", "inputHash": "27a6fb24ab72bcb63d2c14db21d3c45d9bca8fccf872a8d0d70a671ca3d6db52", "model": "deepseek-v4-pro"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We introduce SCHEMEARENA, a 400-scenario benchmark for scalable scheming stress testing", "canonicalNameSource": "abstract", "decisionReason": "The benchmark offers a released dataset, reproducible agent running and scoring scripts, and a public code repository.", "evaluationEvidence": {"reuse": "We release the benchmark, code, and monitor at: https://github.com/launchnlp/SchemeArena.", "scoring": "A trajectory is labeled scheming if the judge finds evidence of a misaligned goal and covert reasoning or covert action, using the rule in _is_scheming() inside evaluate_scheming.py.", "task": "SchemeArena, a 400-scenario benchmark for scalable scheming stress testing, constructed through a factorized scenario synthesis framework spanning diverse safety-relevant tool domains, instrumental goals, oversight conditions, and pressure mechanisms."}, "model": "deepseek-v4-pro", "reviewedAt": "2026-09-10T00:52:28.558291Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates scheming in LLM agents across 400 factorized scenarios, with a monitor that scores covert misaligned behavior along five rubric dimensions.", "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 this end, we introduce SCHEMEARENA, a 400-scenario benchmark for scalable scheming stress testing, constructed through a factorized scenario synthesis framework spanning diverse safety-relevant tool domains, instrumental goals, oversight conditions, and pressure mechanisms."}, "familyId": "bmf_635d7f5ae94f", "firstSeenAt": "2026-09-09", "heat": null, "id": "bm_schemearena-factorized-stress-testing-of-s_9000c40f", "indexedAt": "2026-09-09T06:18:29.794093Z", "industrySectors": [], "links": {"code": "https://github.com/launchnlp/SchemeArena", "data": null, "paper": "https://arxiv.org/abs/2609.08126", "pdf": "https://arxiv.org/pdf/2609.08126", "project": null, "report": "https://arxiv.org/abs/2609.08126"}, "metrics": [], "motivation": "We study scheming in LLM agents, in which agents covertly pursue misaligned goals.", "name": "SchemeArena", "oneLine": "Evaluates scheming in LLM agents across 400 factorized scenarios, with a monitor that scores covert misaligned behavior along five rubric dimensions.", "paperTitle": "SchemeArena: Factorized Stress Testing of Scheming in LLM Agents", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.08126", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-09T06:18:29.794093Z"}, "publications": [], "publishers": [{"name": "LaunchNLP", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/launchnlp/SchemeArena"}], "ranking": {"30d": {"confidence": "Low", "coverage": 0.55, "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": 0.0, "value": 0}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.55, "rank": 134, "score": 23}, "method": "attention-ranking-v13", "rank": 134, "score": 23}, "90d": {"confidence": "Low", "coverage": 0.55, "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": 0.0, "value": 0}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.55, "rank": 329, "score": 23}, "method": "attention-ranking-v13", "rank": 329, "score": 23}, "today": {"confidence": "Low", "coverage": 0.25, "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": 1, "historyDate": "2026-09-08", "rank": null, "score": null, "windowDays": 1}, "level": {"components": {"githubStars": {"normalization": "percentile", "normalized": 0.2857142857142857, "percentile": 0.2857142857142857, "value": 0}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "llmAttentionForecast": {"maxContribution": 0.25, "percentile": 0.55, "role": "bonus", "value": 55}}, "confidence": "Low", "coverage": 0.25, "rank": 6, "score": 46}, "method": "attention-ranking-v13", "rank": 6, "score": 46}}, "readiness": "Runnable", "recognitionConfidence": 0.7, "relation": "introduces", "releasedAt": "2026-09-08", "source": {"authors": ["Jie Ruan", "Inderjeet Nair", "Amy Liu", "Muhammad Khalifa", "Yusheng Zhou", "Lu Wang"], "categories": ["cs.AI", "cs.CL"], "id": "2609.08126", "title": "SchemeArena: Factorized Stress Testing of Scheming in LLM Agents", "type": "arxiv", "url": "https://arxiv.org/abs/2609.08126"}, "sourceUpdatedAt": "2026-09-08T02:02:51Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "Provides a systematic way to stress-test agent scheming under varied goals, oversight, and pressure, informing safer agent deployment."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
