Fractal basins trap latent reasoning
Summary
Reasoning models exhibit transient chaos and fractal basins as a physical consequence of computational complexity when solving difficult tasks.
Training triggers a reasoning bifurcation.
As weights change, incorrect solutions lose stability and become saddles, causing trajectories to transition from smooth convergence to circuitous routes near saddle points.
incorrect solutions lose stability and become saddles
Core variables drive transient chaos.
While simple variables are resolved by substitution, only core variables requiring multi-step Gaussian elimination produce positive finite-time Lyapunov exponents.
Fractal basins emerge across architectures.
The relationship between fractality and task difficulty holds across diverse models, including the Equilibrium Reasoning model on Sudoku and the Parcae model on Countdown.
fractal basins emerge on difficult problems
The FPRM model shows fractality.
During the Maze task, the FPRM model generates a fractal characterized by an uncertainty exponent of approximately 0.853.
Convergence represents system dissipation.
The tendency of reasoning models to converge to fixed answers likely represents a form of dissipation within the dynamical system.
Read the original ↗Save to your Sponge
Summarized with Sponge. Quotes are the author’s own words.