paper-research-assistant

Paper Research Assistant - Research Paper Study and Reproduction Assistant

Core Workflow

1. Paper Reception and Parsing

• Receive the paper PDF file path or arXiv/journal link provided by the user
• Use scripts/parse_paper.py to extract paper metadata (title, authors, abstract, keywords)
• Identify the paper type: theoretical research/experimental research/survey/method paper

2. In-Depth Study and Report Generation

• Extract core contributions (usually located in the last paragraph of the Introduction or in the Conclusion)
• Identify the methodological framework (model architecture, algorithm flow, key formulas)
• Organize the experimental configuration (datasets, baseline methods, evaluation metrics, hyperparameters)
• Generate a structured study report (using references/report_template.md）

3. Resource Collection

• Search for official code repositories (GitHub, GitLab, project homepage)
• Find accompanying datasets (HuggingFace, Kaggle, data sources mentioned in the paper)
• Verify resource availability and licenses

4. Reproduction Code Generation

• Generate a code skeleton based on the paper's method description (PyTorch/TensorFlow)
• Implement the core algorithm modules
• Configure the training loop and evaluation process
• Generate runnable experiment scripts

5. Experimental Plan Design

• List environment dependencies (Python version, key libraries)
• Design comparative experiments (ablation experiments, baseline comparisons)
• Configure the hyperparameter search space
• Estimate computational resource requirements

Script Usage

parse_paper.py

python scripts/parse_paper.py --pdf /path/to/paper.pdf --output /tmp/paper_metadata.json

Extract structured paper metadata

generate_report.py

python scripts/generate_report.py --metadata /tmp/paper_metadata.json --template references/report_template.md --output /tmp/research_report.md

Generate a study report

scaffold_code.py

python scripts/scaffold_code.py --paper-json /tmp/paper_metadata.json --framework pytorch --output-dir /tmp/repo

Generate a reproduction code skeleton

Reference Documents

• references/report_template.md - Standard template for study reports
• references/code_style.md - Reproduction code specifications
• references/experiment_design.md - Experimental design guide

Output Specifications

Study Report Structure

# 论文研读报告

## 基本信息
- 标题：
- 作者/机构：
- 发表 venue：
- 日期：

## 核心贡献
1. ...
2. ...

## 方法论
- 问题定义：
- 核心思路：
- 关键公式：

## 实验配置
- 数据集：
- 基线方法：
- 评估指标：
- 超参数：

## 复现可行性
- 官方代码：[有/无] [链接]
- 数据集：[公开/需申请] [链接]
- 计算需求：
- 预计复现难度：[低/中/高]

## 待澄清问题
- ...

Notes

1. PDF parsing limitations: complex formulas may be recognized inaccurately and require manual verification
2. Scope of code reproduction: generates skeleton code; complete implementation requires actual debugging
3. Resource verification: all links must be verified for validity, with the last access time noted
4. License compliance: indicate the license type of the original paper/code

Tool Dependencies

• PyMuPDF / pdfplumber - PDF parsing
• arxiv API - paper metadata query
• GitHub API - code repository search
• HuggingFace API - dataset query