Reasoning models produce fractals on hard problems; nonlinear dynamics probe their thinking

burny_tech · x · 2026-09-09

The Gilpin lab reports that reasoning models produce fractals when asked to solve hard problems, and that nonlinear dynamics can be used to probe the thinking processes of recurrent-depth models on Sudoku, mathematics, and even ARC-AGI (part 1/N).

The post echoes Jascha Sohl-Dickstein's blog "Neural network training makes beautiful fractals": training and fractal generation share a mechanism—repeatedly applying a parameterized function to its own output—and the hyperparameter boundary between training success and failure has gorgeous fractal structure (paper: arXiv:2402.06184). Together they suggest fractal geometry is pervasive in deep learning dynamics.

Related event: Fractal Basins Trap Latent Reasoning: Overthinking Traced to Transient Chaos(14 posts)→

Original post →

More from Research

Research channel →