Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

UGPhysics: A Comprehensive Benchmark for Undergraduate Physics Reasoning with Large Language Models

Xin Xu, Qiyun Xu, Tong Xiao, Tianhao Chen, Yuchen Yan, Jiaxin Zhang, Shizhe Diao, Can Yang, Yang Wang

TL;DR

UGPhysics introduces a comprehensive, bilingual benchmark (EN/ZH) of 5,520 undergraduate-level physics problems across 13 subjects to evaluate LLMs' physics reasoning. It pairs the dataset with MARJ, a two-stage Model-Assistant Rule-based Judgment framework for robust answer judging, validated against human judgments. An extensive evaluation of 31 LLMs shows a best performance of 49.8% (OpenAI-o1-mini), highlighting that physics reasoning remains challenging despite math-strong models and large scales. The work emphasizes the need for physics-centric training data and evaluation tools, and provides data and code to foster future advances in AI for physics reasoning.

Abstract

Large language models (LLMs) have demonstrated remarkable capabilities in solving complex reasoning tasks, particularly in mathematics. However, the domain of physics reasoning presents unique challenges that have received significantly less attention. Existing benchmarks often fall short in evaluating LLMs' abilities on the breadth and depth of undergraduate-level physics, underscoring the need for a comprehensive evaluation. To fill this gap, we introduce UGPhysics, a large-scale and comprehensive benchmark specifically designed to evaluate UnderGraduate-level Physics (UGPhysics) reasoning with LLMs. UGPhysics includes 5,520 undergraduate-level physics problems in both English and Chinese, covering 13 subjects with seven different answer types and four distinct physics reasoning skills, all rigorously screened for data leakage. Additionally, we develop a Model-Assistant Rule-based Judgment (MARJ) pipeline specifically tailored for assessing answer correctness of physics problems, ensuring accurate evaluation. Our evaluation of 31 leading LLMs shows that the highest overall accuracy, 49.8% (achieved by OpenAI-o1-mini), emphasizes the necessity for models with stronger physics reasoning skills, beyond math abilities. We hope UGPhysics, along with MARJ, will drive future advancements in AI for physics reasoning. Codes and data are available at https://github.com/YangLabHKUST/UGPhysics .

UGPhysics: A Comprehensive Benchmark for Undergraduate Physics Reasoning with Large Language Models

TL;DR

UGPhysics introduces a comprehensive, bilingual benchmark (EN/ZH) of 5,520 undergraduate-level physics problems across 13 subjects to evaluate LLMs' physics reasoning. It pairs the dataset with MARJ, a two-stage Model-Assistant Rule-based Judgment framework for robust answer judging, validated against human judgments. An extensive evaluation of 31 LLMs shows a best performance of 49.8% (OpenAI-o1-mini), highlighting that physics reasoning remains challenging despite math-strong models and large scales. The work emphasizes the need for physics-centric training data and evaluation tools, and provides data and code to foster future advances in AI for physics reasoning.

Abstract

Large language models (LLMs) have demonstrated remarkable capabilities in solving complex reasoning tasks, particularly in mathematics. However, the domain of physics reasoning presents unique challenges that have received significantly less attention. Existing benchmarks often fall short in evaluating LLMs' abilities on the breadth and depth of undergraduate-level physics, underscoring the need for a comprehensive evaluation. To fill this gap, we introduce UGPhysics, a large-scale and comprehensive benchmark specifically designed to evaluate UnderGraduate-level Physics (UGPhysics) reasoning with LLMs. UGPhysics includes 5,520 undergraduate-level physics problems in both English and Chinese, covering 13 subjects with seven different answer types and four distinct physics reasoning skills, all rigorously screened for data leakage. Additionally, we develop a Model-Assistant Rule-based Judgment (MARJ) pipeline specifically tailored for assessing answer correctness of physics problems, ensuring accurate evaluation. Our evaluation of 31 leading LLMs shows that the highest overall accuracy, 49.8% (achieved by OpenAI-o1-mini), emphasizes the necessity for models with stronger physics reasoning skills, beyond math abilities. We hope UGPhysics, along with MARJ, will drive future advancements in AI for physics reasoning. Codes and data are available at https://github.com/YangLabHKUST/UGPhysics .

Paper Structure

This paper contains 32 sections, 5 figures, 23 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. The UGPhysics Benchmark
  4. UGPhysics and MARJ Overview
  5. UGPhysics Creation
  6. MARJ Evaluation Framework
  7. Experiments
  8. Experimental Setup
  9. Main Results
  10. Analysis
  11. Fine-grained Analysis
  12. Reliability of Evaluation
  13. Error Analysis
  14. About Data Leakage
  15. Compare to Existing Benchmarks
  16. ...and 17 more sections

Figures (5)

  • Figure 1: An overall illustration of UGPhysics. The top part represents the hierarchical physics domains and subjects. The bottom part showcases one concrete example.
  • Figure 2: The distribution of overall accuracy across subjects, and physics reasoning skills. (a) The overall accuracy of different subjects averaged across 8 strong LLMs listed in Figure (b). Each bar consists of several segments with colors indicating their corresponding reasoning skills. (b) The overall accuracy of reasoning skills, averaged across all subjects. Only 8 strong LLMs are included for brevity. "KR": Knowledge Recall; "LA": Laws Application; "MD": Math Derivation; "PA": Practical Application; "OT": Others.
  • Figure 3: Performance in different languages, sorted by the difference of EN - ZH.
  • Figure 4: Distribution of Error Types of OpenAI-o1-mini
  • Figure 5: Word Cloud of Topics in UGPhysics