# Equilibrium reasoners use learned attractors to scale test-time compute and lift Sudoku-Extreme accuracy from 2.6% to 99%

> A new AI framework called Equilibrium Reasoners boosted Sudoku accuracy from 2.6% to over 99% by teaching models to converge toward stable solution states. It allocates more compute to harder problems automatically, without external verification.

- Type: AI news
- Page: https://completeaitraining.com/news/equilibrium-reasoners-use-learned-attractors-to-scale-test/
- Published: 2026-05-22
- Updated: 2026-05-23
- Author: Jeroen Erne
- Jobs: Science and research
- Reading time: 2 min

### Researchers Show How AI Models Learn to Solve Complex Problems Through Stable States

A new framework called Equilibrium Reasoners (EqR) explains how iterative AI models generalize beyond memorization by learning internal dynamical systems that converge toward correct answers. The approach enables dramatic accuracy gains on reasoning tasks by scaling test-time compute - the computational work performed when solving a specific problem - without requiring external verification systems.

The core finding: models that learn to recognize stable solution states can allocate compute adaptively, spending minimal effort on simple problems while dedicating massive computational resources to hard ones.

#### How Learned Attractors Drive Reasoning

Rather than treating generalization as a fixed property of model architecture, EqR frames it as a dynamic process. Models learn what researchers call "attractor landscapes" - internal maps where stable points correspond to valid solutions. During inference, the model's computations iteratively move toward these stable states, much like a ball rolling downhill toward equilibrium.

This mechanism explains why scaling test-time compute improves performance. More iterations allow the model to converge more fully toward solution-aligned states.

#### Adaptive Compute Allocation Enables Extreme Scaling

The framework allocates computational effort based on task difficulty. Simple problems converge in a handful of iterations. Complex reasoning tasks benefit from massive scaling - the researchers tested computations equivalent to 40,000 neural network layers on a challenging Sudoku benchmark.

The results: accuracy jumped from 2.6% using a standard feedforward model to over 99% with learned attractors and scaled test-time compute. The same approach works across different problem types without task-specific tuning.

This adaptive allocation matters for deployment. Real systems can't afford unlimited compute per problem, but EqR's internal mechanisms determine when additional iterations will help.

#### Implications for Iterative Reasoning Models

The attractor-based view provides a mechanistic explanation for how iterative latent models achieve scalable reasoning. Rather than treating test-time scaling as an empirical hack, it grounds the approach in dynamical systems theory.

For researchers working on [generative AI and LLMs](https://completeaitraining.com/tag/generative-ai-and-llm/), this framework offers a new lens for understanding when and why scaling computation helps. It also suggests which architectural choices support learning effective attractor structures.

The work was presented at ICML 2026. For those building or studying reasoning systems, understanding attractor-based generalization could inform approaches to test-time scaling and model design. [AI research courses](https://completeaitraining.com/tag/research/) covering iterative models and dynamical systems provide relevant background for deeper engagement with this framework.

---

This is the Markdown form of https://completeaitraining.com/news/equilibrium-reasoners-use-learned-attractors-to-scale-test/. The whole site is mapped at https://completeaitraining.com/llms.txt.