Research group finds reasoning models produce fractals when solving hard problems

wgilpin0 · x · 2026-09-09

The team discovered that reasoning models produce fractals when solving hard problems. They parametrize problem + weights as a dynamical system with solutions as fixed points, varying each replicate's initial latent state across a random 2D slice. Nonlinear dynamics lets them probe recurrent depth models on Sudoku, math, and even ARC-AGI, and explains 'overthinking': similar prompts can yield 10x differences in reasoning length.

Related event: Fractal basins and transient chaos explain why reasoning models overthink(13 posts)→

Original post →

More from Research

Research channel →