EqR: Scalable Reasoning via Latent Attractors

In this AI Research Roundup episode, Alex discusses the paper: 'Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning' This paper introduces Equilibrium Reasoners (EqR), a novel framework that scales test-time compute by learning task-conditioned attractors in a latent dynamical system. Unlike traditional methods, EqR scales internal dynamics by increasing iteration depth and aggregating stochastic trajectories without relying on external verifiers. This approach allows neural networks to adaptively allocate compute based on task difficulty, resolving simple cases quickly while scaling massively for harder ones. Empirically, unrolling EqR up to the equivalent of 40,000 layers boosts accuracy on Sudoku-Extreme from 2.6% to over 99%. Paper URL: https://arxiv.org/pdf/2605.21488 #AI #MachineLearning #DeepLearning #EquilibriumReasoners #TestTimeCompute #ReasoningModels #LatentDynamics

Видео EqR: Scalable Reasoning via Latent Attractors канала AI Research Roundup