Energy-Based Reasoners: 385x more efficient inference

312 points · 55 comments on HN · read original →

Points and comments are a snapshot, not live.

Energy-Based Reasoners achieve 99.9% accuracy on Sudoku with 345x less compute than next-best.

EBRs learn an energy function over latent reasoning states, using gradient descent to refine and stop. With only 1.8M parameters, they reach 99.9% accuracy on Sudoku-Extreme using 345x fewer FLOPs than EqR. On N-Queens and Graph Coloring, they achieve state-of-the-art coverage with near-100% single-sample accuracy. The model scores candidate predictions, guides refinement via gradient, and uses gradient norm for adaptive stopping. Exploration-based training shapes the energy landscape over broader states.

What commenters are saying

Several commenters note the paper works on constrained puzzle domains, not general language reasoning. One commenter explains the energy mechanism concretely: a learned matrix assigns higher energy to conflicting digits, with probability inversely proportional to energy. Another links to Logical Intelligence, a company promoting EBMs, which some call a scam. A debate emerges on whether excluding code execution from LLM demos limits the definition of reasoning. One commenter clarifies that EBMs differ from traditional models by dropping normalization, with P(x) ∝ e^(−E(x)).