Energy-Based Reasoners: Recurrent Reasoners Should Be Energy-Based Models

William Xingxu Chen · hf · 2026-11-18

The paper proposes Energy-Based Reasoners: modeling looped latent reasoning as an energy-based model. A learned energy function's gradient refines reasoning states step by step, while its value serves as a stopping signal telling the model when to exit the loop. Exploration-based training further improves both performance and efficiency.

Related event: Energy-Based Reasoners: Learning When to Stop in Recurrent Reasoning(3 posts)→

Original post →

More from Research

Research channel →