Researchers find reasoning models produce fractals when solving hard problems like ARC-AGI

ShahabBakht · x · 2026-09-24

wgilpin0's group discovered that reasoning models produce fractal structures when asked to solve hard problems. Using nonlinear dynamics, they probe the thinking processes of recurrent-depth models on Sudoku, mathematics, and even ARC-AGI.

Quoting commenter notes the work echoes Jascha Sohl-Dickstein's mini exploration paper on chaotic frontiers in neural networks; they are working on "chaotic training" methods and propose a workshop on the theme.

Original post →

More from Research

Research channel →