Definition
Plain language
A grid-maze benchmark used to test small recursive-reasoning models.
As stated in the literature
A hard variant of grid-maze pathfinding used as a structured-reasoning benchmark for tiny recursive reasoners (HRM, TRM) and attractor-style models.
Why it matters: It's a clean structured-reasoning task that separates models that genuinely plan from those that mostly pattern-match.
For example, a tiny recursive model is asked to find a path through a dense grid maze where many corridors are dead-ends and naive search blows up.
Heard on the show
“Sudoku-Extreme and Maze-Hard.”Episode 041 — When the Iteration Teaches the Model to Skip the Iteration