Definition
Plain language
A grid of extremely hard Sudoku puzzles used to test small reasoning models.
As stated in the literature
A high-difficulty Sudoku benchmark used to evaluate tiny recursive reasoners (HRM, TRM) and attractor-style architectures on algorithmic problem solving.
Why it matters: It gives a clean, algorithmic test of reasoning ability that doesn't depend on world knowledge, useful for studying small specialized architectures.
For example, a 7-million-parameter recursive reasoner can be evaluated on Sudoku-Extreme to see if tiny models can match the puzzle-solving of much larger ones.
Heard on the show
“TRM at seven million parameters gets about seventy-five percent on Sudoku-Extreme.”Episode 041 — When the Iteration Teaches the Model to Skip the Iteration