Your highest accuracy

[Chase08765]

What’s your highest accuracy on problems like Sudoku, ARC, or maze solving when using TRM?

[Ilakiancs]

on structured reasoning tasks, trm reaches around 40% on arc-agi-1 and over 77% on sudoku and maze benchmarks. even with only 5–7m params, the recursive latent state update cycles allow deep multi-step reasoning without large-scale model overhead. cost efficiency still needs tuning, but the reasoning stability is impressive for this scale.

[olivkoch]

FYI I've built a branch that provides cleaner experimentation and training on a simple problem (Sudoku 4x4). This allows anyone to debug/understand/experiment with TRM rapidly.

https://github.com/olivkoch/TinyRecursiveModels

[olivkoch]

I reach 87% exact accuracy on the validation set with Sudoku Extreme with the implementation above (it matches the number in the paper). I am getting 75% exact val accuracy on Maze Hard.

https://github.com/olivkoch/nano-trm