Energy-Based Reasoners: Your Recurrent Reasoner Should Be an Energy-Based Model

energy-based-reasoners.github.io2 min read

Researchers introduce Energy-Based Reasoners, a novel class of models that use an explicit energy function over latent reasoning states to guide looped updates via gradient descent. The energy scores states, while its gradient provides a natural stopping criterion and improvement direction. Across multiple benchmarks, Energy-Based Reasoners match or outperform prior recurrent reasoning models while using significantly less inference compute. Open article →

More on this topic