Energy-based reasoner adds built-in progress signal to looped reasoning models
Looped reasoning models reuse a small network repeatedly to gain depth without extra parameters, but they lack a way to judge their latent states, decide when to stop, or pick the best candidate answer. A paper from Duke, UIUC and Harvard proposes Energy-Based Reasoners, which learn an energy function that scores how compatible a latent state is with the input. The same function drives the update as gradient descent, so the model halts when the gradient becomes small and can select the lowest-energy trajectory among several. The authors report that their 1.8M-parameter model reaches 99.9% accuracy on Sudoku-Extreme using 345 times less inference compute than the next-best compared model. Training includes an exploration component meant to shape the energy landscape beyond the correct solutions and to handle puzzles with multiple valid answers.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.



Discussion (0)
Log in to join the discussion and vote.
Log in