On September 4, 2026, researchers posted “Fractal basins trap latent reasoning” (arXiv:2609.04963), a study of how frontier reasoning models behave while solving hard problems. Testing diverse leading reasoning models across Sudoku, maze solving, visual puzzles, and mathematical logic, the authors show these models are dynamical systems whose latent reasoning trajectories have fractal basin boundaries - and that the fractality of those basins increases directly with task difficulty.
The proposed mechanism is that models become trapped near saddle points in their reasoning trajectory that represent nearly-correct attempted solutions. As a problem gets harder, the geometry around those saddle points grows more convoluted, which is what produces the long, meandering reasoning traces observed when frontier models attack theorem proving, coding, or multi-step planning. The paper frames this as transient chaos: the model is not stuck by accident, it is navigating a genuinely fractal decision landscape.
The paper’s central claim is that extended test-time compute on hard problems is not mainly a product of RL training incentives or verbose habits picked up during post-training. It is closer to an inevitable consequence of a problem’s underlying computational complexity, expressed as chaotic dynamics in the reasoning process. The authors argue reasoning traces deserve study as their own class of dynamical system, with tools borrowed from chaos theory rather than pure NLP evaluation.
For anyone building or budgeting agentic systems, this reframes “how long will the model think” as a property of problem structure rather than a setting that can reliably be prompted or fine-tuned away. It does not show that longer reasoning produces correct answers, only that compute usage on hard instances has a specific, predictable geometric signature - a useful diagnostic, not yet a fix for slow or runaway reasoning traces.