TinyRecursiveModels is a recursive training framework for small neural networks designed to solve complex logical tasks. It functions as a parameter-efficient model trainer and a reasoning dataset generator, enabling the optimization of models that refine their answers through iterative reasoning steps.
The framework differentiates itself by utilizing latent-state recursive refinement, where the model maintains and updates an internal hidden representation to improve prediction accuracy over multiple sequential steps. It also includes tools for generating structured training and evaluation datasets based on logical puzzles and maze solving.
The system covers hardware-accelerated training loops and parameter-efficient network design to reduce computational overhead while maintaining reasoning capabilities.