{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2605.24305", "kind": "paper", "url": "https://arxiv.org/abs/2605.24305"}], "authors": ["Noel Thomas"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2605.24305", "source_url": "https://arxiv.org/abs/2605.24305"}, "categories": ["General AI", "Language & Knowledge", "Reasoning"], "description": {"en": "ChaosBench-Logic v2 evaluates LLM logical reasoning over dynamical systems with 40,886 questions across 165 systems, 27 first-order logic predicates, and 78 axiom edges. The CARE protocol measures calibration and adversarial robustness, reporting metrics such as MCC."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2605.24305", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2605.24305", "title": "ChaosBench-Logic"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Mohamed bin Zayed University of Artificial Intelligence"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2605.24305", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2605.24305", "source_url": "https://openalex.org/W7162403078"}, "key": "claire-radar:2605.24305", "modality": null, "name": "ChaosBench-Logic", "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_chaosbench-logic_c5cd44be", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "4dd6b4439e408bad06a4bdfdd45a4c371c6ceea1d443ef7687b7a2a1b8803e12", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2605.24305"}, "publisher": null, "released": "2026-05-23", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2605.24305", "source_url": "https://arxiv.org/abs/2605.24305"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2605-24305", "source": "claire_radar", "source_benchmark_id": "2605.24305", "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/2605.24305", "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": ["Reasoning"], "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": "9c846a7cbecc884598e92cb84a3b54515adcf6bc3b71d939178af3da7363ec4f", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "ChaosBench-Logic v2 evaluates LLM logical reasoning over dynamical systems with 40,886 questions across 165 systems, 27 first-order logic predicates, and 78 axiom edges. The CARE protocol measures calibration and adversarial robustness, reporting metrics such as MCC.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We present ChaosBench-Logic v2, a 40,886-question benchmark over 165 dynamical systems with 27 FOL predicates and 78 axiom edges, together with CARE (Calibration- and Adversarial-Robust Evaluation), a protocol that surfaces these pathologies."}, "familyId": "bmf_173654d46d35", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_chaosbench-logic_c5cd44be", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2605.24305", "paper": "https://arxiv.org/abs/2605.24305", "pdf": "https://arxiv.org/pdf/2605.24305", "project": null, "report": "https://arxiv.org/abs/2605.24305"}, "metrics": [], "motivation": "Standard accuracy on binary reasoning benchmarks hides critical failure modes: prior collapse, inconsistency under paraphrase, and inability to reason about parameter-dependent dynamics.", "name": "ChaosBench-Logic", "oneLine": "ChaosBench-Logic v2 evaluates LLM logical reasoning over dynamical systems with 40,886 questions across 165 systems, 27 first-order logic predicates, and 78 axiom edges. The CARE protocol measures calibration and adversarial robustness, reporting metrics such as MCC.", "paperTitle": "ChaosBench-Logic v2: Evaluating LLM Logical Reasoning over Dynamical Systems at Scale", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2605.24305", "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": "2026-05-23", "source": {"authors": ["Noel Thomas"], "categories": ["cs.LG", "cs.AI"], "id": "2605.24305", "title": "ChaosBench-Logic v2: Evaluating LLM Logical Reasoning over Dynamical Systems at Scale", "type": "arxiv", "url": "https://arxiv.org/abs/2605.24305"}, "sourceUpdatedAt": "2026-05-23T00:34:17Z", "topics": ["Reasoning"], "venueAttempts": [], "whyItMatters": "Binary accuracy on reasoning benchmarks hides failures like prior collapse and inconsistency under paraphrase. This benchmark surfaces these pathologies and quantifies reasoning about parameter-dependent dynamics, guiding improvements in LLM reasoning capabilities."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
