| cff-version: 1.2.0 |
| message: "If you use this benchmark in your research, please cite it as below." |
| type: software |
| title: "ChaosBench-Logic: A Benchmark for Evaluating Large Language Models on Complex Reasoning about Dynamical Systems" |
| version: 2.0.0 |
| date-released: 2026-02-18 |
| url: "https://github.com/11NOel11/ChaosBench-Logic" |
| repository-code: "https://github.com/11NOel11/ChaosBench-Logic" |
| license: MIT |
| authors: |
| - family-names: "Thomas" |
| given-names: "Noel" |
| affiliation: "Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)" |
| keywords: |
| - "large language models" |
| - "benchmark" |
| - "dynamical systems" |
| - "chaos theory" |
| - "reasoning" |
| - "evaluation" |
| - "LLM" |
| - "GPT-4" |
| - "Claude" |
| - "Gemini" |
| - "LLaMA" |
| abstract: > |
| ChaosBench-Logic is a comprehensive benchmark for evaluating Large Language |
| Models (LLMs) on complex reasoning tasks involving chaotic and non-chaotic |
| dynamical systems. The benchmark consists of 40,886 v2 questions plus 621 |
| archived v1 questions (41,507 total) spanning 165 dynamical systems (30 core |
| + 135 from dysts library) across 10 task families: atomic, multi-hop, |
| indicator diagnostics, regime transition, FOL inference, cross-indicator, |
| extended systems, consistency paraphrase, adversarial, and perturbation |
| robustness. The ontology covers 27 predicates with first-order logic axioms |
| enabling up to 4-hop reasoning chains. The codebase supports multiple |
| state-of-the-art LLMs (GPT-4, Claude-3.5, Gemini-2.5, LLaMA-3 70B) in |
| both zero-shot and chain-of-thought modes, providing comprehensive metrics |
| including overall accuracy, task-specific breakdowns, FOL violation |
| detection, and bias analysis. |
| references: |
| - type: article |
| authors: |
| - family-names: "Thomas" |
| given-names: "Noel" |
| title: "ChaosBench-Logic: Benchmarking Large Language Models on Complex Reasoning about Dynamical Systems" |
| year: 2025 |
| journal: "arXiv preprint" |
|
|
