---
title: "UT Austin Finds Fractal Chaos Explains Why AI Reasoning Costs Spike 10x"
subtitle: "A new paper shows recurrent-depth reasoning models behave like chaotic dynamical systems, where hard problems create fractal basins that trap thinking near wrong answers."
url: "https://alphasignal.ai/news/ut-austin-finds-fractal-chaos-explains-why-ai-reasoning-costs-spike-10x"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-09-08T19:01:24.009Z"
updated: "2026-09-08T19:03:59.026Z"
type: "Paper"
topics:
  - "Reasoning"
  - "Benchmarks"
subtopics:
  - "Test Time Compute"
  - "Math Reasoning"
upvotes: 4696
source_url: "https://arxiv.org/abs/2609.04963"
comments: 0
truncated: true
---

# UT Austin Finds Fractal Chaos Explains Why AI Reasoning Costs Spike 10x

> A new paper shows recurrent-depth reasoning models behave like chaotic dynamical systems, where hard problems create fractal basins that trap thinking near wrong answers.

A team from UT Austin has borrowed tools from nonlinear physics to explain something practitioners keep noticing but nobody could pin down: why two nearly identical prompts can produce reasoning traces that differ by ten times in length and token cost. Their answer is that [recurrent-depth reasoning models](https://arxiv.org/abs/2609.04963) are dynamical systems with **fractal basins of attraction**, and harder problems make the fractality worse.

The setup treats reasoning as an optimizer. A reasoning model and problem instance jointly define an autonomous discrete-time dynamical system acting on a latent state, and the latent state after each loop can be decoded as an intermediate solution representation. The final answer is a fixed point that trajectories converge to. What the authors probe is how the starting point of that trajectory affects convergence time.

## Coloring the map of thought

Pick a model and a problem. The latent state for modern reasoning models is typically initialized by sampling a random vector. The authors continuously vary the initial latent state across a random 2D slice, then color each pixel by how many loops it took to reach the answer.

The maps look like Mandelbrot sets. Adjacent starting points take wildly different routes and iteration counts to reach the same correct answer. This is the physical signature of

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [Fractal basins trap latent reasoning (arXiv 2609.04963)](https://arxiv.org/abs/2609.04963)
- [Fractal basins trap latent reasoning (HTML)](https://arxiv.org/html/2609.04963v1)
- [GilpinLab/loopscape](https://github.com/GilpinLab/loopscape)
- [William Gilpin thread on the paper](https://twitter.com/wgilpin0/status/2097354911017316736)