Doctoral Speaking Skills Talk - Zhengyang Geng

Many reasoning models can improve by spending more computation at test time, but it remains unclear what makes this scaling effective. This talk introduces Equilibrium Reasoners (EqR), which view reasoning as convergence toward task-conditioned attractors: stable latent states that correspond to correct solutions. EqR scales computation through additional iterations and multiple stochastic trajectories, without external verifiers or task-specific priors. Across challenging reasoning tasks, improved performance closely tracks convergence toward solution-aligned attractors, offering a new framework for understanding and scaling latent reasoning.