EqR-data

This repository contains the datasets used in the paper Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning.

The datasets are designed to evaluate scalable test-time reasoning in iterative latent models, specifically focused on learning task-conditioned attractors.

Dataset Details

The repository includes data for two main reasoning tasks:

• Sudoku-Extreme: A set of challenging Sudoku puzzles designed to test the limits of iterative reasoning models, following the setup from HRM.
• Maze-Unique: A dataset of 30x30 mazes where each maze has a unique solution path and specific length constraints.

Links

• Paper: Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning
• GitHub: locuslab/eqr
• Project Page: X (Twitter) Thread

Usage

The datasets can be downloaded using the scripts provided in the official GitHub repository:

git clone https://github.com/locuslab/eqr
cd eqr
bash scripts/download_artifacts.sh

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}
}