Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning

Researchers introduced Equilibrium Reasoners, a framework where iterative latent updates converge to task-specific attractors.

By scaling depth and breadth at test time, the method can turn weak feedforward baselines into much higher-accuracy reasoners on hard benchmarks.

This provides a mechanistic explanation for why iterative models generalize and shows that large test-time compute can boost performance without external verifiers.