Source-linked AI summary
Fractal basins trap latent reasoning
Jeffrey Lai, Anthony Bao, John Quinn, William Gilpin
TL;DR
The paper investigates how reasoning-model dynamics can be characterized through basin complexity and fractality across tasks, architectures, and scales. It extracts latent trajectories, measures settling-time basins and their boundaries, and finds difficult problems produce fractal basins, including true and slim fractals across models.
Problem
The study addresses how to quantify basin complexity and determine whether reasoning-model basins exhibit scale-free fractality across tasks and architectures.
Method
The authors extract latent states and decoded outputs at every reasoning loop, map settling times over two-dimensional initial-condition slices, and measure basin entropy, boundary entropy, uncertainty exponents, Lyapunov indicators, and latent trajectory geometry.
Results
Difficult problems produce fractal basins across alternative reasoning architectures, while models differ between true fractals and scale-dependent slim fractals.
Takeaways & Limitations
Reasoning-model basin geometry depends on task difficulty, architecture, and resolution, with convergence tendencies likely contributing dissipation and doubly transient chaos.
Abstract
from arXiv · showhide
Reasoning allows artificial intelligence models to revisit and correct their mistakes, enabling recent frontier advances in mathematical theorem solving, software engineering, and autonomous task planning. Reasoning models are widely observed to reason for longer on harder tasks, but the general mechanism responsible for these slowdowns is unknown. Here, we show that reasoning models exhibit transient chaos, a physical consequence of the computational complexity of difficult tasks. As a consequence, we show that diverse leading reasoning models are dynamical systems with fractal basins, with fractality increasing with task difficulty across diverse tasks like Sudoku and maze solving, visual puzzles, and mathematical logic. We show that transient chaos emerges due to reasoning becoming trapped for extended durations near saddle points, which we show correspond to nearly-correct attempted solutions of the underlying problem. Our results show that reasoning slowdowns are an inevitable consequence of problem hardness in modern artificial intelligence models, and establish reasoning traces as a rich new class of dynamical system.
Appendix Appendix B: Supplementary Methods
The supplementary methods define the recurrent-model dynamics, task sampling, basin-map construction, and complexity measurements used to analyze reasoning trajectories. They also specify the Lyapunov and latent-trajectory analyses used to identify route boundaries and saddle behavior.
- Models and experiments: Experiments use publicly released recurrent models whose latent states and decoded outputs are available at every reasoning loop.EqR dynamics are made deterministic by disabling injected noise and fixing the loop budget; FPRM halts at a residual threshold subject to an iteration cap.
- Basin maps: For each initial condition, the decoded solution state is tracked until it stops changing, producing a settling-time field that serves as the basin map.Trajectories still changing at the loop cap are marked as errors.
- Task instances and difficulty: Task instances span Sudoku, maze, and Countdown-style reasoning settings with difficulty defined by benchmark ratings, shortest-path length, or generated maze structure.Generated mazes use 30 × 30 boards, with solution paths ranging from 101–169 steps and optional route multiplicity of 2^k.
- Basin complexity: Basin entropy and related measures quantify local variation in settling times, while slices with insufficient convergence or excessive capped trajectories are excluded from population statistics.The analysis computes basin entropy, boundary entropy, and uncertainty exponent across at least 300 valid slices per model-task pair.
- Route boundaries: The fast Lyapunov indicator identifies initial conditions whose trajectories separate strongly from neighbors before reconverging, marking saddle-like boundaries between solution routes.Large λF is computed from the maximum local deviation between each pixel’s decoded state and those of its initial nearest neighbors.
- Trajectory geometry: Principal components analysis is applied to pooled latent trajectories to study their geometry relative to the final latent solution.This analysis examines full latent dynamics rather than only decoded solution-space trajectories.
Appendix Appendix C: Replicate basin experiments with alternative reasoning models
The authors replicate basin analyses with alternative architectures on Maze and Countdown tasks. Difficult problems again produce fractal basins, while basin appearance varies with the optimization algorithm.
- Replication: Alternative Equilibrium Reasoning and encoder-only looped-transformer architectures produce fractal basins on difficult Maze and Countdown problems.The replication was designed to test whether fractal basins depended on particular task-model pairings.
- Difficulty dependence: Fractality increases with task difficulty across the alternative architectures and tasks.The reported relationship matches the pattern observed with the other models.
- Architecture dependence: Basin appearance and structure change across architectures, likely because they reflect the particular optimization algorithm used to solve each task.
Appendix Appendix D: Metrics for basin complexity
The appendix compares three basin-complexity metrics computed on random two-dimensional slices through task instances. These metrics are strongly correlated across puzzle instances and tasks.
- Metrics: Basin entropy measures average convergence-time variation within finite-resolution neighborhoods of initial conditions.
- Metrics: Boundary basin entropy restricts the same variation measure to initial conditions that straddle multiple basins.
- Metrics: The uncertainty exponent α measures how the fraction f(ε) of outcomes changed by perturbations of size ε scales as f(ε) ∼ ε^α.With sampled-space dimension D = 2, α = D − d_b relates the exponent to the basin-boundary box-counting dimension.
- Cross-metric comparison: The three complexity measures are broadly strongly correlated across different puzzle instances and tasks.The relationship follows from the connection between uncertainty-exponent scaling and basin-boundary dimension.
Appendix Appendix E: Measuring fractality across scales
The appendix tests whether fractality persists across scales by zooming into basin and Lyapunov fields and measuring uncertainty-exponent scaling. Models can produce true or slim fractals, with convergence tendencies likely contributing dissipation.
- Scale visualization: Progressive magnification examines convergence-time and fast-Lyapunov-indicator fields to visualize basin structure across scales.The supplementary figures show zoom sequences for Sudoku, Maze, and Countdown examples.
- Scale measurement: The uncertainty exponent quantifies how outcome variation within a fixed-radius ε neighborhood changes with ε, testing for a power-law signature of scale-free fractality.A fixed power-law relationship between uncertainty and radius supports the standard fractal interpretation.
- Fractal types: Some reasoning models produce true fractals, whereas others produce slim fractals whose fractal dimension decreases with resolution.Slim fractals have scale-dependent structure and typically arise when saddle-like sets decay over time in dissipative systems.
- Dynamical interpretation: The authors expect many reasoning models to exhibit some form of doubly transient chaos because convergence to fixed answers likely acts like dissipation.This is presented as an expectation rather than a universal empirical finding.