jacobfa1/attractor-140m

Solve the Loop: Attractor Models for Language and Reasoning

This repository contains the Attractor-140M model presented in Solve the Loop: Attractor Models for Language and Reasoning.

Project Page | GitHub | Paper

Introduction

Attractor Models are a family of models that use a backbone module to propose output embeddings and an attractor module to refine them by solving for a fixed point through implicit differentiation. This architecture allows training memory to remain constant relative to effective depth and enables iterations to be chosen adaptively. In language modeling, Attractor Models deliver a Pareto improvement over standard Transformers and stable looped models across sizes.

Sample Usage

To use this model, you first need to install the package from the official repository:

pip install -e .

Then you can initialize the model as follows:

from attractor.models.attractor import Attractor, AttractorConfig

config = AttractorConfig.from_name("attractor-small-140m")
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}
}