Reasoning models' fractal basins are smooth and naturalistic, resembling optical caustics or turbulent transient lifetimes, despite the discrete puzzles being solved

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Hardness-induced fractal basins are ubiquitous in modern reasoning models, emerging even without hand-designed task-specific solvers, across highly distinct architectures and tasksclaim0.863
  The paper's generalizing claim distinguishing its contribution from prior narrow demonstrations (e.g., Ercsey-Ravasz & Toroczkai's SAT solver).
• Fractal basins form because reasoning trajectories become trapped near weakly-unstable saddle points for extended durationsclaim0.840
  The paper's core mechanistic claim connecting saddle dynamics to basin fractality.
• Fractal basins and transient chaos are the generic manner in which computational complexity manifests in analogue systemsclaim0.835
  Generalizing interpretation connecting this paper's AI findings to physical analogue computation broadly.
• Most reasoning models are expected to exhibit some form of doubly transient chaos, because properties like convergence-to-fixed-answers act as a dissipation-like mechanismhypothesis0.810
  Forward-looking predictive claim about reasoning models generally, based on the analogy to damped physical systems.
• Fractal basins and their scaling with difficulty replicate on alternative architecture-task pairings (EqR-Maze, encoder-only looped transformer on Countdown)finding0.802
  Replication check ruling out that fractal basins arise only from one model-task pairing.
• "Here, we show that reasoning models exhibit transient chaos, a physical consequence of the computational complexity of difficult tasks."quote0.801
  Load-bearing abstract sentence stating the paper's central discovery.
• Fractal Basinsconcept0.777
  Self-similar convergence-time basins the paper finds in reasoning models' latent initial-condition space, growing with task difficulty.
• Basin entropy strongly correlates with number of reasoning iterations required to converge, across Sudoku, mazes, visual puzzles, and mathematical logicfinding0.776
  Core empirical result establishing fractality scales with task difficulty across four tasks/architectures.