license: mit
pipeline_tag: text-generation
Solve the Loop: Attractor Models for Language and Reasoning
Attractor Models are a family of models that use a backbone module to propose output embeddings, followed by an attractor module that refines them by solving for a fixed point using implicit differentiation. This architecture allows for iterative refinement with constant training memory and adaptive inference-time computation.
Project Page | Paper (arXiv:2605.12466) | GitHub
Introduction
Attractor Models offer a promising alternative to feed-forward computation by iteratively refining latent representations. In language modeling, Attractor Models deliver a Pareto improvement over standard Transformers, improving perplexity and downstream accuracy while reducing training cost. This repository contains the Attractor-370M model.
Sample Usage
To use this model, you need to install the attractor package from the official repository:
git clone https://github.com/jacobfa/Attractor
cd Attractor
pip install -e .
Then, you can construct the model in Python:
from attractor.models.attractor import Attractor, AttractorConfig
# Loading the configuration for the 370M model
config = AttractorConfig.from_name("attractor-medium-370m")
model = config.construct_model()
Citation
@article{feinashley2026attractor,
title={Solve the Loop: Attractor Models for Language and Reasoning},
author={Fein-Ashley, Jacob and Rashidinejad, Paria},
year={2026},
url={https://arxiv.org/abs/2605.12466}
}