Energy-Based Reasoners: gradient descent on latent states for adaptive reasoning

kalyan_kpl · x · 2026-11-22

A new paper introduces Energy-Based Reasoners (EBRs), which learn an energy function over latent reasoning states and refine them via gradient descent toward lower-energy configurations. The same energy function guides reasoning updates, ranks candidate solutions, and enables adaptive stopping via gradient convergence—solving the lack of explicit progress measures in existing recurrent reasoners. It also proposes State-based Explorative Modeling (SXM) to explore multiple latent trajectories and learn diverse valid solution modes. Evaluated on tasks including Sudoku-Ex.

Original post →

More from Research

Research channel →