Less Is More: Recursive Reasoning with Tiny Networks

How a 7‑million‑parameter model teaches giant AIs a lesson in thinking

Solving mazes and Sudoku can be done efficiently as well with smaller models. Image created with DALL-E.

Artificial intelligence has taught itself to write emails, caption pictures and summarize long reports, yet it still stumbles on the kind of bite‑sized logic puzzles a bright teenager loves: Sudokus, twisty mazes, and the infamous ARC‑AGI pattern chal…