Researchers from Duke University, UIUC, and Harvard University introduced Energy-Based Reasoners (EBRs), formulating looped latent reasoning as gradient descent on a single learned energy function that simultaneously guides state refinement, triggers adaptive stopping, and ranks candidate predictions without an external verifier. With 1.8M parameters, EBRs reached 99.9% accuracy on Sudoku-Extreme while using 345x less inference compute than the next-best baseline.
Energy-Based Reasoners (EBRs)
Also known as: EBR · EBRs · Energy-Based Reasoning
Plain English
An alternative reasoning architecture that treats multi-step problem solving as sliding downhill on a learned energy landscape rather than generating long chains of text tokens. The same energy score tells the model how to refine its state, when to stop thinking, and which candidate answer to keep.
Technical Overview
Energy-Based Reasoners (EBRs) learn a scalar energy function over latent reasoning states and replace learned recurrent transition rules with gradient descent on that energy. Because the energy value itself scores candidate solutions without an external verifier and its gradient guides iterative refinement and early stopping, EBRs can dramatically reduce test-time inference compute compared with autoregressive Transformer or recurrent reasoning baselines.
Chronological Entries (1)
Direct Relations (2)
- Alternative toTransformer
- UsesEvals & Benchmarks
Referenced By (0)
No incoming references yet.