This repository contains the official implementation of our Buffer of Thoughts (BoT) framework. Affiliation: Peking University, UC Berkeley, Stanford University

• Release initial code of BoT, supporting GPT-4 and Llama3-70B [2024.6.6]
• Update the code for smaller LLMs (e.g., Llama3-8B) [2024.6.24]
• Release meta-buffer and buffer-manager
• Extending BoT to more applications

We introduce BoT, a novel and versatile thought-augmented reasoning approach designed to enhance the accuracy, efficiency, and robustness of large language models (LLMs). Specifically, we propose a meta-buffer to store a series of high-level thoughts, referred to as thought-templates, distilled from problem-solving processes across various tasks. For each problem, we retrieve a relevant thought-template and adaptively instantiate it with specific reasoning structures to conduct efficient reasoning. To ensure scalability and stability, we also propose a buffer-manager to dynamically update the meta-buffer, thus enhancing its capacity as more tasks are solved. We conduct extensive experiments on 10 challenging reasoning-intensive tasks, achieving significant performance improvements over previous state-of-the-art (SOTA) methods: 11% on Game of 24, 20% on Geometric Shapes, and 51% on Checkmate-in-One. Further analysis demonstrates the superior generalization ability and robustness of our BoT, while requiring only 12% of the cost of multi-query prompting methods (e.g., tree/graph of thoughts) on average. Notably, we find that our Llama3-8B + BoT has the potential to surpass Llama3-70B model.

Overview of our BoT

For now, we release our demo version of BoT based on three different benchmarks:

• The Game of 24 from Yao et al., 2023
• Checkmate-in-One from the BIG-Bench suite (BIG-Bench authors, 2023)
• Word Sorting from BIG-Bench Hard (Suzgun et al., 2023; BIG-Bench authors, 2023)

For each task, we choose one thought template sampled from our meta-buffer library. Stay tuned for our complete meta-buffer library update!

First, set up the environment:

git clone https://github.com/YangLing0818/buffer-of-thought-llm
cd buffer-of-thought-llm
conda create -n BoT python==3.9 
pip install -r requirements.txt

Our BoT is easy to use. Just run:

python run_benchmarks.py --task_name 'gameof24' --api_key 'input your API key here if you want to use GPT-4' --model_id 'the model ID of GPT-4 or the path to your local LLM'

Here, --task_name could be one of gameof24, checkmate, wordsorting.

The --api_key is required if you want to use GPT-series; if not, you can skip it.

The --model_id should be the model ID of GPT-series like gpt-4o, gpt-4-turbo, or the path to your local LLM if you do not set --api_key.

The data for these three tasks are located in the /benchmarks directory.

The results generated during the experiment are stored in the /test_results directory.

Run the command below to validate the test results of our BoT:

python validate_results.py --task_name 'gameof24' --test_path 'The path to the .jsonl file you want to validate'

This will print out the accuracy of the selected task on your relevant .jsonl file.

@article{yang2024buffer,
  title={Buffer of Thoughts: Thought-Augmented Reasoning with Large Language Models},
  author={Yang, Ling and Yu, Zhaochen and Zhang, Tianjun and Cao, Shiyi and Xu, Minkai and Zhang, Wentao and Gonzalez, Joseph E and Cui, Bin},
  journal={arXiv preprint arXiv:2406.04271},
  year={2024}
}