README.md

metadata

license: mit
pretty_name: PRBench
language:
  - en
tags:
  - physics
  - benchmark
  - agents
  - scientific-reproduction
  - paper-reproduction
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: tasks.jsonl

PRBench: Physics Paper Reproduction Benchmark

PRBench evaluates whether an AI agent can read a physics paper and reproduce its computational results from scratch: understand the method, implement the simulation or calculation without the paper's original code, run it, and produce figure/table data that a model judge compares against withheld reference results.

This repository contains the public task specifications for 31 tasks spanning plasma physics, nuclear and particle theory, lattice QCD, strong-field and attosecond physics, quantum optics, trapped-ion quantum computing, condensed matter, heavy-ion hydrodynamics, atomic structure, astrophysics, and computational electromagnetics.

What is included

Each task directory holds exactly what the solving agent is given:

File	Purpose
task.yaml	Task configuration: paper info, expected output files, Docker environment, grading rubric (full and code-only modes)
instruction.md	The instruction sent to the agent: target figures/tables, required code files, CSV output formats, and (for most tasks) banned libraries
<paper>.md	The paper, converted to Markdown (the agent's only view of the paper)
images, images/, images1/	Paper figures copied into the agent workspace (only for tasks that have them)
eosq.dat	Input data file required by task_kolb_2003 (declared in input_files)

tasks.jsonl is a one-row-per-task index (paper metadata, full instruction text, expected outputs, environment, and rubric weights) that drives the dataset viewer.

What is withheld

To keep the benchmark usable for evaluation, grading references are not published:

• metadata.md: ground-truth formulas and methodology used by the judge
• data/: reference CSV outputs
• reproduce/ / reproduction/: reference implementations
• original source archives and PDFs

Fields in task.yaml such as metadata_file, ground_truth_data_dir, and ground_truth_code_dir refer to these withheld files.

Evaluation protocol

An evaluation run copies the paper Markdown, figures, and input files into an isolated Docker workspace and sends instruction.md to the agent. The agent must write:

1. reproduction/ANALYSIS.md: its analysis of the paper's methodology,
2. the code files listed under expected_outputs.code,
3. the CSV files listed under expected_outputs.data.

Only after the agent finishes is the ground truth placed in the workspace, and a model judge scores the submission on the rubric dimensions in task.yaml (typically methodology understanding, code correctness, data accuracy, and completeness, with data accuracy weighted most heavily). In code-only mode, the code_only.instruction_suffix is appended to the instruction, the agent writes code without running it, and the code_only.grading rubric applies.

Loading

from datasets import load_dataset
tasks = load_dataset("Rise-AGI/PRBench", split="test")

# or fetch the full task directories
from huggingface_hub import snapshot_download
path = snapshot_download("Rise-AGI/PRBench", repo_type="dataset")

Tasks

Task ID	Paper	Authors	Year	Reference
task_bude_2021	Accelerating simulations of electromagnetic waves in hot, magnetized fusion plasmas	R H S Budé, D Van Eester, J van Dijk, R J E Jaspers and A B Smolders	2021	10.1088/1361-6587/abd619
task_davighi_2024	Topological Portal to the Dark Sector	Joe Davighi, Admir Greljo, Nudžeim Selimović	2024	arXiv:2401.09528
task_dimitrov_2000	Chirality of Nuclear Rotation	S. Dimitrov, T. Frauendorf, F. Dönau	2000	10.1103/PhysRevLett.84.5732
task_dong_2025	Enhanced nonlinear Hall effect by Cooper pairs near the superconducting phase transition	Zi-Hao Dong, Hui Yang, Yi Zhang	2025	10.1103/PhysRevB.111.155120
task_fu_2001	Classical collisional trajectories as the source of strong-field double ionization of helium in the knee regime	L. B. Fu, J. Liu, J. Chen, S. G. Chen	2001	10.1103/PhysRevA.63.043416
task_fu_2008	Energetic-Particle-Induced Geodesic Acoustic Mode	Guoyong Fu	2008	10.1103/PhysRevLett.101.185002
task_fukushima_2004	Chiral effective model with the Polyakov loop	Kenji Fukushima	2004	10.1016/j.physletb.2004.04.027
task_gralla_2018	Observational Signature of High Spin at the Event Horizon Telescope	Samuel E. Gralla, Alexandru Lupsasca, Andrew Strominger	2018	10.1093/mnras/sty039
task_harris_1989	Lasers without Inversion: Interference of Lifetime-Broadened Resonances	S. E. Harris	1989	10.1103/PhysRevLett.62.1033
task_harris_1999	Nonlinear Optics at Low Light Levels	S. E. Harris and Lene Vestergaard Hau	1999	10.1103/PhysRevLett.82.4611
task_ishikawa_1982	The glueball mass spectrum in QCD: first results of a lattice Monte Carlo calculation	K. Ishikawa, M. Teper, G. Schierholz	1982	—
task_kolb_2003	Hydrodynamic description of ultrarelativistic heavy-ion collisions	Peter F. Kolb and Ulrich Heinz	2003	nucl-th/0305084v2
task_li_2014	Classical-Quantum Correspondence for Above-Threshold Ionization	Min Li, Ji-Wei Geng, Hong Liu, Yongkai Deng, Chengyin Wu, Liang-You Peng, Qihuang Gong, Yunquan Liu	2014	10.1103/PhysRevLett.112.113002
task_lin_2020	Bootstraps to Strings: Solving Random Matrix Models with Positivity	Henry W. Lin	2020	arXiv:2002.08387
task_liu_2025	Scattering Entanglement Entropy and Its Implications for Electroweak Phase Transitions	Jia Liu, Masanori Tanaka, Xiao-Ping Wang, Jing-Jun Zhang, Zifan Zheng	2025	arXiv:2505.06001v2
task_luscher_1991	Signatures of unstable particles in finite volume	Martin Luscher	1991	10.1016/0550-3213(91)90584-K
task_monmayrant_2010	A newcomer's guide to ultrashort pulse shaping and characterization	Antoine Monmayrant, Sébastien Weber, Béatrice Chatel	2010	10.1088/0953-4075/43/10/103001
task_parke_1986	Amplitude for n-Gluon Scattering	Stephen J. Parke and T. R. Taylor	1986	10.1103/PhysRevLett.56.2459
task_ratti_2006	Phases of QCD: Lattice thermodynamics and a field theoretical model	Claudia Ratti, Michael A. Thaler, Wolfram Weise	2006	10.1103/PhysRevD.73.014019
task_roothaan_1960	Self-Consistent Field Theory for Atomic Systems	Clemens C. J. Roothaan	1960	10.1103/RevModPhys.32.186
task_rosenbluth_1998	Poloidal Flow Driven by Ion-Temperature-Gradient Turbulence in Tokamaks	M. N. Rosenbluth, F. L. Hinton	1998	10.1103/PhysRevLett.80.724
task_shvetsov_2016	Semiclassical two-step model for strong-field ionization	N. I. Shvetsov-Shilovski et al.	2016	10.1103/PhysRevA.94.013415
task_vautherin_1972	Hartree-Fock Calculations with Skyrme's Interaction. I. Spherical Nuclei	D. Vautherin and D. M. Brink	1972	10.1103/PhysRevC.5.626
task_wang_2015	Quantum Computation under Micromotion in a Planar Ion Crystal	S.-T. Wang, C. Shen, L.-M. Duan	2015	10.1038/srep08555
task_white_1993	Density-matrix algorithms for quantum renormalization groups	Steven R. White	1993	10.1103/PhysRevB.48.10345
task_wu_2013	Time-domain Perspective on Autler-Townes Splitting in Attosecond Transient Absorption	Mengxi Wu, Shaohao Chen, Mette B. Gaarde, Kenneth J. Schafer	2013	10.1103/PhysRevA.88.043416
task_wu_2018	Noise analysis for high-fidelity quantum entangling gates in an anharmonic linear Paul trap	Yukai Wu, Sheng-Tao Wang, L.-M. Duan	2018	10.1103/PhysRevA.97.062325
task_wu_2021	High-fidelity entangling gates in a three-dimensional ion crystal under micromotion	Y.-K. Wu, Z.-D. Liu, W.-D. Zhao, L.-M. Duan	2021	10.1103/PhysRevA.103.022419
task_xie_2025	Developing a linear fluid plasma model with accurate kinetic Bernstein waves: A first step	Huasheng Xie	2025	10.1063/5.0274495
task_xue_2023	Generalized Quantum Measurements on a Higher-Dimensional System via Quantum Walks	Peng Xue	2023	10.1103/PhysRevLett.131.150803
task_yee_1966	Numerical Solution of Initial Boundary Value Problems Involving Maxwell's Equations in Isotropic Media	Kane S. Yee	1966	10.1109/TAP.1966.1138693

Notes

• Paper Markdown was produced by automatic PDF conversion. Some image links point to external hosts or to figures that are not included.
• Paper texts and figures remain the copyright of their original authors and publishers and are included solely for research evaluation. The MIT license applies to the task specifications (instructions, configurations, and index).