DISCOVERPHYSICS
收藏资源简介:
DISCOVERPHYSICS是由普林斯顿大学等研究机构构建的交互式基准数据集,旨在评估大型语言模型在非标准物理世界中的科学发现能力。该数据集包含22个精心设计的模拟世界,每个世界通过N体模拟器动态生成粒子轨迹数据,数据规模灵活可调,涵盖短程指数屏蔽力、分数阶拉普拉斯算子及隐藏粒子物种等多样化物理定律。数据集的创建过程基于可控的仿真环境,允许智能体主动设计实验并观察原始轨迹,以迭代方式推断底层物理规律。该数据集主要应用于人工智能与科学发现交叉领域,旨在解决模型从噪声观测中识别相关特征、构建机制模型并最终发现非常规运动方程的核心挑战,从而推动对模型长程推理与概念理解能力的深入测评。
DISCOVERPHYSICS is an interactive benchmark dataset developed by Princeton University and other research institutions, aiming to evaluate the scientific discovery capabilities of large language models (LLMs) in non-standard physical worlds. This dataset comprises 22 meticulously designed simulated worlds, each dynamically generating particle trajectory data via an N-body simulator with flexibly adjustable data scale, covering diverse physical laws such as short-range exponential screening forces, fractional Laplacian operators, and hidden particle species. The dataset is built on a controllable simulation environment, which allows AI Agents to actively design experiments, observe raw trajectories, and iteratively infer underlying physical laws. Primarily applied in the interdisciplinary field of artificial intelligence and scientific discovery, this dataset targets addressing core challenges including models identifying relevant features from noisy observations, constructing mechanistic models, and finally discovering unconventional equations of motion, so as to facilitate in-depth evaluations of models' long-range reasoning and conceptual understanding capabilities.




