System-1.x: Learning to Balance Fast and Slow Planning with Language Models

Paper

Swarnadeep Saha, Archiki Prasad, Justin Chih-Yao Chen, Peter Hase, Elias Stengel-Eskin, and Mohit Bansal

Comparative Overview of System-1, System-2, and System-1.x Planning

System-1.x

Installation

This repository is tested on Python 3.9.16.
You should install this repository on a virtual environment. All dependencies can be installed as follows:

pip install -r requirements.txt

Datasets

The maze and blocksworld data are included in data folder. Each sample is a json, describing the start state, goal state, and the system 1/2 traces on which the models will be trained.

Experiments on Maze Navigation

To run experiments on maze navigation, check out the scripts inside scripts/maze.

For example, the command to train a system-1 model is:

bash scrips/maze/train_sys1.sh

Then you can evaluate the model using the following command:

bash scrips/maze/eval_sys1.sh

To train and evaluate our final System-1.5 model with sub-goal decomposition, first train a System-1 model and a System-2 model. Then execute the following commands:

bash scrips/maze/train_sys1.5_sg.sh
bash scrips/maze/eval_sys1.5_sg.sh

You can train any System-1.x model by just setting the value of x. Similarly, you can alter the search algorithm by passing --search_algo dfs/bfs.

Experiments on Blocksworld

Experiments on Blocksworld follow a similar pattern. Check out the scripts inside scrips/blocksworld.

Citation

@article{saha2023system-1.x,
  title={System-1.x: Learning to Balance Fast and Slow Planning with Language Models},
  author={Saha, Swarnadeep and Prasad, Archiki and Chen, Justin Chih-Yao and Hase, Peter and Stengel-Eskin, Elias and Bansal, Mohit},
  journal={arXiv preprint arXiv:2407.14414},
  year={2024}
}