Update README.md
Browse files
README.md
CHANGED
|
@@ -10,117 +10,163 @@ tags:
|
|
| 10 |
|
| 11 |
# PhysReason: A Comprehensive Benchmark towards Physics-Based Reasoning
|
| 12 |
|
| 13 |
-
[
|
| 14 |
-
[
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
## Key Features
|
| 24 |
-
|
| 25 |
-
-
|
| 26 |
-
-
|
| 27 |
-
-
|
| 28 |
-
-
|
| 29 |
-
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
-
|
| 38 |
-
|
| 39 |
-
-
|
| 40 |
-
|
| 41 |
-
- Classification
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
|
| 11 |
# PhysReason: A Comprehensive Benchmark towards Physics-Based Reasoning
|
| 12 |
|
| 13 |
+
[](https://arxiv.org/abs/2502.12054)
|
| 14 |
+
[](https://huggingface.co/datasets/zhibei1204/PhysReason)
|
| 15 |
+
[](https://dxzxy12138.github.io/PhysReason/)
|
| 16 |
+
|
| 17 |
+
> **PhysReason is accepted by ACL-2025-main**
|
| 18 |
+
|
| 19 |
+
## 📋 Overview
|
| 20 |
+
|
| 21 |
+
PhysReason is a comprehensive physics-based reasoning benchmark consisting of **1,200 physics problems** spanning multiple domains, with a focus on both knowledge-based (25%) and reasoning-based (75%) questions. This benchmark addresses the critical gap in evaluating large language models' capabilities in physics-based reasoning, which requires applying physics theorems and constraints in complex problem-solving scenarios.
|
| 22 |
+
|
| 23 |
+
## ✨ Key Features
|
| 24 |
+
|
| 25 |
+
- **📊 Dataset Size**: 1,200 carefully curated physics problems
|
| 26 |
+
- **🎯 Problem Types**: Strategic mix of knowledge-based (25%) and reasoning-based (75%) questions
|
| 27 |
+
- **📚 Theorem Coverage**: Comprehensive coverage of 147 physics theorems
|
| 28 |
+
- **🎨 Visual Content**: 81% of problems include diagrams and visual elements
|
| 29 |
+
- **📈 Difficulty Levels**: Four distinct levels - Knowledge, Easy, Medium, Hard
|
| 30 |
+
- **🔄 Step-by-step Solutions**: Average of 8.1 solution steps per problem (15.6 for hard problems)
|
| 31 |
+
- **🌍 Multi-modal**: Supports both text and image inputs
|
| 32 |
+
|
| 33 |
+
## 🔧 Data Collection
|
| 34 |
+
|
| 35 |
+
Our rigorous data collection process ensures high-quality, challenging problems:
|
| 36 |
+
|
| 37 |
+
- **📖 Sources**: Global college entrance exams and international physics competitions
|
| 38 |
+
- **⚙️ Process**: Standardized using MinerU framework for consistent formatting
|
| 39 |
+
- **✅ Quality Control**: Two-phase translation process with expert verification
|
| 40 |
+
- **🔍 Filtering**: Systematically excluded easily searchable problems to prevent data leakage
|
| 41 |
+
- **📊 Classification**: Difficulty levels based on solving time and theorem complexity analysis
|
| 42 |
+
|
| 43 |
+
## 📊 Benchmark Comparison
|
| 44 |
+
|
| 45 |
+
| Benchmark | Multi-modal | Size | Knowledge | Question Type | Avg. T | Step-by-step | Avg. T | Avg. S |
|
| 46 |
+
|----------------|-------------|------|-----------|---------------|--------|--------------|--------|--------|
|
| 47 |
+
| JEEBench | ❌ | 123 | CEE | OE,MC | 169.7 | - | - | - |
|
| 48 |
+
| MMLU-Pro | ❌ | 1299 | COL | MC | 52.1 | - | - | - |
|
| 49 |
+
| GPQA | ❌ | 227 | PH.D. | OE | 111.4 | ❌ | 197.2 | 3.6 |
|
| 50 |
+
| SciEval | ❌ | 1657 | - | OE,MC | 154.5 | - | - | - |
|
| 51 |
+
| SciBench | ✅ | 295 | COL | OE | 80.5 | ❌ | 315.9 | 2.8 |
|
| 52 |
+
| MMMU | ✅ | 443 | COL | OE,MC | 53.8 | - | - | - |
|
| 53 |
+
| ScienceQA | ✅ | 617 | K1-K12 | MC | 13.3 | ❌ | 63.0 | 2.4 |
|
| 54 |
+
| OlympiadBench | ✅ | 2334 | COMP | OE | 222.0 | ❌ | 199.8 | 3.7 |
|
| 55 |
+
| EMMA | ✅ | 156 | - | MC | 109.5 | - | - | - |
|
| 56 |
+
| **Ours-Knowledge** | ✅ | 300 | CEE+COMP | OE | 163.7 | ✅ | 196.5 | 3.3 |
|
| 57 |
+
| **Ours-Easy** | ✅ | 300 | CEE+COMP | OE | 171.2 | ✅ | 241.5 | 5.0 |
|
| 58 |
+
| **Ours-Medium** | ✅ | 300 | CEE+COMP | OE | 229.2 | ✅ | 391.3 | 8.4 |
|
| 59 |
+
| **Ours-Hard** | ✅ | 300 | CEE+COMP | OE | 340.9 | ✅ | 936.1 | 15.6 |
|
| 60 |
+
| **Ours-Full** | ✅ | 1200 | CEE+COMP | OE | 226.3 | ✅ | 441.3 | 8.1 |
|
| 61 |
+
|
| 62 |
+
## 🔍 Evaluation Framework
|
| 63 |
+
|
| 64 |
+
We introduce the **Physics Solution Auto Scoring (PSAS)** framework with two complementary evaluation approaches:
|
| 65 |
+
|
| 66 |
+
### PSAS-A (Answer Level Evaluation)
|
| 67 |
+
- **Sub-question Assessment**: Evaluates answers for each sub-question independently
|
| 68 |
+
- **LLM-based Extraction**: Uses advanced language models for answer extraction
|
| 69 |
+
- **Semantic Verification**: Ensures semantic consistency between extracted and ground truth answers
|
| 70 |
+
- **Weighted Scoring**: Considers solution step lengths as weights for different sub-questions
|
| 71 |
+
|
| 72 |
+
### PSAS-S (Step Level Evaluation)
|
| 73 |
+
Provides detailed step-by-step assessment through four phases:
|
| 74 |
+
1. **Data Extraction**: Parses model responses and reference solutions
|
| 75 |
+
2. **Scoring**: Evaluates correctness of each reasoning step
|
| 76 |
+
3. **First Error Detection**: Identifies where models first deviate from correct reasoning
|
| 77 |
+
4. **Error Analysis**: Classifies error types into four key bottlenecks:
|
| 78 |
+
- Physics Theorem Application
|
| 79 |
+
- Physics Process Understanding
|
| 80 |
+
- Calculation
|
| 81 |
+
- Physics Condition Analysis
|
| 82 |
+
|
| 83 |
+
## 🚀 Usage
|
| 84 |
+
|
| 85 |
+
### Core Evaluation Files
|
| 86 |
+
- `answer_evaluation_with_ds_ch_prompt.py`: Answer-level evaluation using Chinese prompts
|
| 87 |
+
- `answer_evaluation_with_ds_en_prompt.py`: Answer-level evaluation using English prompts
|
| 88 |
+
- `format_result_ds.py`: Optimizes unstable outputs into stable, consistent formats
|
| 89 |
+
- `step_evaluation_with_ds_ch_prompt.py`: Step-level evaluation using Chinese prompts
|
| 90 |
+
- `step_evaluation_with_ds_en_prompt.py`: Step-level evaluation using English prompts
|
| 91 |
+
|
| 92 |
+
## 📈 Experimental Results
|
| 93 |
+
|
| 94 |
+
### Non-O-like Models Performance
|
| 95 |
+
|
| 96 |
+
| Model | Input | Knowledge | Easy | Medium | Hard | Avg. |
|
| 97 |
+
|-------------------|-------|-------------|-------------|-------------|-------------|-------------|
|
| 98 |
+
| Qwen2VL-72B | Q, I | 41.92/62.47 | 24.04/45.26 | 15.97/36.13 | 4.83/24.23 | 16.96/42.88 |
|
| 99 |
+
| InternVL2.5-78B | Q, I | 28.34/64.71 | 24.16/50.69 | 17.72/38.56 | 9.71/25.95 | 19.98/45.89 |
|
| 100 |
+
| GPT-4o | Q, I | 50.71/65.82 | 33.87/51.98 | 22.73/42.36 | 11.03/24.71 | 29.58/47.23 |
|
| 101 |
+
| Deepseek-V3-671B | Q, IC | 55.86/66.14 | 40.06/52.77 | 26.63/44.02 | 13.73/26.87 | 34.07/48.42 |
|
| 102 |
+
| Claude-3.5-Sonnet | Q, I | 54.14/66.45 | 41.35/55.85 | 28.14/44.86 | 15.11/28.51 | 34.69/49.88 |
|
| 103 |
+
| Gemini-2.0-Flash | Q, I | 65.08/75.04 | 54.84/68.60 | 39.79/55.67 | 21.99/38.39 | 45.20/60.40 |
|
| 104 |
+
| Gemini-2.0-Pro | Q, I | 67.99/79.01 | 55.43/71.47 | 44.29/57.74 | 23.81/42.66 | 47.88/62.74 |
|
| 105 |
+
|
| 106 |
+
### O-like Models Performance
|
| 107 |
+
|
| 108 |
+
| Model | Input | Knowledge | Easy | Medium | Hard | Avg. |
|
| 109 |
+
|------------------------------------|-------|-------------|-------------|-------------|-------------|-------------|
|
| 110 |
+
| o1-mini | Q, IC | 53.90/65.74 | 35.21/52.26 | 22.24/40.19 | 10.61/26.80 | 30.49/47.18 |
|
| 111 |
+
| QvQ-72B | Q, I | 62.44/70.92 | 53.74/64.65 | 28.18/54.88 | 14.30/36.47 | 32.67/57.66 |
|
| 112 |
+
| Gemini-2.0-Flash-Thinking-1206 | Q, I | 65.35/77.20 | 51.89/67.49 | 44.43/58.95 | 27.14/45.48 | 47.20/63.07 |
|
| 113 |
+
| QwQ-32B | Q, IC | 62.03/76.28 | 54.92/71.08 | 43.64/62.14 | 22.99/42.19 | 45.89/63.87 |
|
| 114 |
+
| GLM-Zero | Q, IC | 64.95/80.36 | 54.11/71.54 | 41.32/63.67 | 23.04/47.46 | 46.52/65.76 |
|
| 115 |
+
| o3-mini-high | Q, IC | 70.67/83.61 | 67.20/81.95 | 45.31/64.57 | 30.12/47.23 | 53.32/69.34 |
|
| 116 |
+
| Gemini-2.0-Flash-Thinking-0121 | Q, I | 73.44/84.15 | 63.17/75.94 | 50.41/66.60 | 31.90/48.47 | 54.73/69.73 |
|
| 117 |
+
| **Deepseek-R1** | Q, IC | **75.11/85.91** | **65.08/79.81** | **54.84/72.02** | **31.95/51.50** | **56.75/73.26** |
|
| 118 |
+
|
| 119 |
+
### PhysReason-mini Results
|
| 120 |
+
|
| 121 |
+
| Model | K. | E. | M. | H. | Avg. |
|
| 122 |
+
|------------------------------------|-------|-------|-------|-------|-------|
|
| 123 |
+
| o1-mini | 54.80 | 30.33 | 15.41 | 7.92 | 27.11 |
|
| 124 |
+
| QvQ-72B | 51.17 | 37.10 | 29.83 | 22.13 | 35.06 |
|
| 125 |
+
| QwQ-32B | 64.40 | 50.07 | 38.88 | 27.45 | 45.20 |
|
| 126 |
+
| Gemini-2.0-Flash-Thinking-1206 | 71.47 | 49.97 | 36.83 | 22.97 | 45.42 |
|
| 127 |
+
| GLM-Zero | 72.70 | 50.17 | 43.42 | 24.70 | 47.75 |
|
| 128 |
+
| o1 | 72.47 | 53.37 | 49.31 | 25.32 | 50.12 |
|
| 129 |
+
| o3-mini-high | 71.10 | 63.20 | 47.02 | 31.93 | 53.31 |
|
| 130 |
+
| Gemini-2.0-Flash-Thinking-0121 | 76.33 | 56.87 | 51.85 | 32.61 | 54.42 |
|
| 131 |
+
| **Deepseek-R1** | **85.17** | **60.77** | **47.24** | **33.23** | **56.60** |
|
| 132 |
+
|
| 133 |
+
## 🔑 Key Findings
|
| 134 |
+
|
| 135 |
+
- **Performance Gap**: Even top-performing models achieve less than 60% on answer-level evaluation
|
| 136 |
+
- **Difficulty Scaling**: Performance drops significantly from knowledge questions (75.11%) to hard problems (31.95%)
|
| 137 |
+
- **O-like Model Advantage**: Models with enhanced reasoning capabilities show superior performance
|
| 138 |
+
- **Multi-modal Benefits**: Visual content significantly enhances model understanding and performance
|
| 139 |
+
- **Four Critical Bottlenecks** identified through step-level evaluation:
|
| 140 |
+
1. **Physics Theorem Application**
|
| 141 |
+
2. **Physics Process Understanding**
|
| 142 |
+
3. **Calculation Accuracy**
|
| 143 |
+
4. **Physics Condition Analysis**
|
| 144 |
+
|
| 145 |
+
## 📝 Citation
|
| 146 |
+
|
| 147 |
+
If you find PhysReason useful in your research, please cite our paper:
|
| 148 |
+
|
| 149 |
+
```bibtex
|
| 150 |
+
@article{zhang2025physreason,
|
| 151 |
+
title={Physreason: A comprehensive benchmark towards physics-based reasoning},
|
| 152 |
+
author={Zhang, Xinyu and Dong, Yuxuan and Wu, Yanrui and Huang, Jiaxing and Jia, Chengyou and Fernando, Basura and Shou, Mike Zheng and Zhang, Lingling and Liu, Jun},
|
| 153 |
+
journal={arXiv preprint arXiv:2502.12054},
|
| 154 |
+
year={2025}
|
| 155 |
+
}
|
| 156 |
+
```
|
| 157 |
+
|
| 158 |
+
## 📄 License
|
| 159 |
+
|
| 160 |
+
This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details.
|
| 161 |
+
|
| 162 |
+
## 📧 Contact
|
| 163 |
+
|
| 164 |
+
We welcome contributions to PhysReason! Please contact us for more details.
|
| 165 |
+
|
| 166 |
+
---
|
| 167 |
+
|
| 168 |
+
**🔗 Quick Links:**
|
| 169 |
+
- [📄 Paper](https://arxiv.org/abs/2502.12054)
|
| 170 |
+
- [🤗 Dataset](https://huggingface.co/datasets/zhibei1204/PhysReason)
|
| 171 |
+
- [🌐 Project Page](https://dxzxy12138.github.io/PhysReason/)
|
| 172 |
+
- [💻 GitHub Repository](https://github.com/dxzxy12138/PhysReason)
|