metadata

license: mit
task_categories:
  - question-answering
language:
  - en
tags:
  - reasoning
  - physics
size_categories:
  - n<1K
arxiv:
  - 2502.15815
configs:
  - config_name: default
    data_files:
      - split: public
        path: data/public-*
dataset_info:
  features:
    - name: problem_id
      dtype: string
    - name: domain
      dtype: string
    - name: difficulty_level
      dtype: string
    - name: problem
      dtype: string
    - name: solution
      dtype: string
    - name: answer
      dtype: string
    - name: code_answer_requirements
      dtype: string
    - name: reference_implementation
      dtype: string
  splits:
    - name: public
      num_bytes: 52743
      num_examples: 10
  download_size: 29845
  dataset_size: 52743

TP Bench – Theoretical Physics Benchmark for AI

TPBench is a curated dataset and evaluation suite designed to measure the reasoning capabilities of AI models in theoretical physics. Our test problems span multiple difficulty levels—from undergraduate to frontier research—and cover topics such as cosmology, high-energy theory, general relativity, and more. By providing a unified framework for problem-solving and auto-verifiable answers, TPBench aims to drive progress in AI-based research assistance for theoretical physics.

Dataset Sources

Paper: https://arxiv.org/abs/2502.15815

Website: https://tpbench.org/

Citation

If you do find our dataset helpful, please cite our paper.

BibTeX:

@misc{chung2025theoreticalphysicsbenchmarktpbench,  
  title={Theoretical Physics Benchmark (TPBench) -- a Dataset and Study of AI Reasoning Capabilities in Theoretical Physics},   
  author={Daniel J. H. Chung and Zhiqi Gao and Yurii Kvasiuk and Tianyi Li and Moritz Münchmeyer and Maja Rudolph and Frederic Sala and Sai Chaitanya Tadepalli},  
  year={2025},  
  eprint={2502.15815},  
  archivePrefix={arXiv},  
  primaryClass={cs.LG},  
  url={https://arxiv.org/abs/2502.15815},   
}