This repository provides partial datasets and the automated code for generating graph problem instances as described in the paper "GTA: Graph Theory Agent and Benchmark for Algorithmic Graph Reasoning with LLMs." Our work introduces GT Bench, a challenging new benchmark featuring 44 diverse graph problem types designed to evaluate the ability of LLMs to perform multi-step algorithmic reasoning. We also propose the Graph Theory Agent (GTA), a novel framework that enhances LLM graph reasoning by employing an adaptive input representation selector and decomposing algorithmic solutions.

Here you will find:

• Code to automatically generate graph problem instances for GT Bench.
• Sample datasets used in our evaluations.

For full details on the methodology, benchmark, agent framework, and experimental results, please refer to our paper.

To set up the environment for running the code, please follow these steps:

conda create -n gta python=3.9
conda activate gta
pip install -r requirements.txt

To use models that require API keys (e.g., OpenAI models, or models hosted on services like DeepInfra), you need to configure your credentials.

1. Create a file named .env in the utils/ directory of this project.
2. Add the necessary API keys and credentials to this file. You only need to add keys for the models or data sources you intend to use.

Here is an example structure for your .env file:

# For OpenAI models
OPENAI_API_KEY=your_openai_api_key
# For models hosted on DeepInfra (if you are using it as an OpenAI-compatible endpoint)
DEEPINFRA_BASE_URL=https://api.deepinfra.com/v1/openai
DEEPINFRA_API_KEY=your_deepinfra_api_key
# Add other API keys as needed for different services
# ANOTHER_SERVICE_API_KEY=your_other_service_key

Make sure the .env file is placed in the correct utils/ directory.

The main scripts provided are:

• graph.py: This script is used for generating the underlying graph structures based on various parameters.
• question.py: This script generates the actual problem instances, which include a natural language question description and the corresponding graph structure (in one of the supported formats).
  • You can control the number of questions generated per problem type using the --questions command-line argument.
  • The scale and types of graphs generated for different tasks can be adjusted by modifying the TASKS configuration (likely a dictionary or list) within the question.py script.
• evaluation.py: This script is used to evaluate the performance of different LLMs on the generated graph problems.
  • You can control which models are evaluated by modifying the models list or variable within the script.
  • You can select the input graph representation types (e.g., Natural Language, Adjacency List, Adjacency Matrix, Structured Language) for evaluation by modifying the data_formats list or variable within the script.

If you use GT Bench, GTA, or the associated code in your research, please cite our paper:

@article{xu2025gta,
  title={GTA: Graph Theory Agent and Benchmark for Algorithmic Graph Reasoning with LLMs},
  author={Xu, Zixiang and Wang, Yanbo and Wang, Chenxi and Gao, Lang and Song, Zirui and Huang, Yue and Chen, Zhaorun and Zhang, Xiangliang and Chen, Xiuying},
  year={2025},
  note={Under review}
}