metadata
license: apache-2.0
pipeline_tag: other
Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning
This repository contains the models presented in the paper Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning.
- Code: GitHub Repository
- Project Page: X/Twitter Thread
Equilibrium Reasoners (EqR) enable test-time scaling without external verifiers or task-specific priors by learning task-conditioned attractors. This approach allows neural networks to adaptively allocate test-time compute based on task difficulty by scaling internal dynamics along two axes: depth (iterations) and breadth (stochastic trajectories).
Citation
@article{huang2026equilibrium,
title={Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning},
author={Huang, Benhao and Geng, Zhengyang and Kolter, Zico},
journal={arXiv preprint arXiv:2605.21488},
year={2026}
}