ReSEED

This code was used to generate the results in "ReSeeding Latent States for Sequential Language Understanding".

Installation

Below are the instructions for installing and running the code on linux using conda.

Create conda environment

conda create -n reseed python=3.10

Activate conda environment

conda activate reseed

Install gym minigrid repository for dataset creation

pip install -e git+https://github.com/StephAO/gym-minigrid.git#egg=minigrid

Install ReSEED

Please note that if you need a non-default pytorch installation, here is a good place to do it.

git clone <TODO>
cd ReSEED
pip install -e .

Keys (Optional)

Fill in necessary keys in reseed/personal_keys for desired functionality. There are three personal keys defined:

WANDB_ENTITY: if specified and account is logged into in the cli, logs experiments to wandb  
OPENAI_API_KEY = used for experiments with openai models (i.e., when calling icl_promting with an openai model)  
ANTHROPIC_API_KEY = used for experiments with anthropic models (i.e., when calling icl_promting with an anthropic model)  

Create Datasets

python -m text_traj_datasets.create_datasets

Usage

python -m contrastive_concepts.main

For scripts to run experiments found in the paper, see scripts/run_n_sample_sweep.sh, scripts/run_ablations.sh, and run_icl.sh.

Notes

NOTE: The [STATE] token described in the paper is defined using the [CLS] token due to huggingface tokenizer nomenclature.

Citation

If you find this code useful, please cite the following paper:

@inproceedings{
aroca-ouellette2025reseeding,
title={ReSeeding Latent States for Sequential Language Understanding},
author={St{\'e}phane Aroca-Ouellette and Katharina von der Wense and Alessandro Roncone},
booktitle={The 2025 Conference on Empirical Methods in Natural Language Processing},
year={2025},
url={https://openreview.net/forum?id=9fd2YtkQk5}
}