introvoyz041/CritPt
收藏资源简介:
CritPt(Complex Research using Integrated Thinking – Physics Test,读作critical point)是首个旨在测试大语言模型在未发表的研究级推理任务上的基准,广泛覆盖现代物理研究领域,包括凝聚态物理、量子物理、原子分子与光学物理、天体物理、统计物理、核物理、高能物理、数学物理、流体动力学、非线性动力学和生物物理。该数据集由来自全球30多个领先机构的50多名物理研究人员(包括资深博士生、博士后、教授和研究科学家)精心设计。CritPt的挑战旨在反映适合物理初级博士生的全面研究项目的深度和复杂性。数据集中的所有问题都是新创建的、在发布前未被大语言模型见过、定义明确具有无歧义的约束和定义、可基于公开知识解决(即不需要机密或私人信息),并且需要深度理解前沿物理研究背景和精确的多步推理与执行。该数据集是包含70个挑战的测试集。每个问题都具有抗猜测性和机器可验证的答案,可通过针对高级物理特定输出格式高度定制的自动评分流程进行评估。
CritPt (Complex Research using Integrated Thinking – Physics Test; reads as critical point) is the first benchmark designed to test LLMs on unpublished, research-level reasoning tasks that broadly covers modern physics research areas, including condensed matter, quantum physics, atomic, molecular & optical physics, astrophysics, statistical physics, nuclear physics, high energy physics, mathematical physics, fluid dynamics, nonlinear dynamics and biophysics. It is crafted by a team of 50+ physics researchers from 30+ leading institutions worldwide, including senior Ph.D. students, postdocs, professors and research scientists. CritPt challenges are designed to reflect the depth and complexity of full-scale research projects suitable for junior Ph.D. students in physics. All problems in CritPt are newly created and unseen by LLMs prior to this release, well-posed with unambiguous constraints and definitions, feasible to solve with the publicly available knowledge (i.e., no confidential or private information is needed), and requiring a deep understanding of the frontier physics research context and precise multi-step reasoning and execution. This dataset is the test set of 70 challenges. Every problem admits a guess-resistant and machine-verifiable answer and can be evaluated by an automated grading pipeline heavily customized for advanced physics-specific output formats.




