ResearchNovember 18, 2026
27
š¶ļø
Energy Model Shatters Sudoku Record
Energy-Based Reasoners solve Sudoku with 345x less inference compute.
#reasoning#energy-based-models#inference-efficiency#recurrent-models#benchmarks

š„ What happened
Researchers at UNC and MIT unveiled Energy-Based Reasoners (EBRs): recurrent models that steer latent states via a learned energy function. Instead of iterating blindly, the model follows the energy gradient downhill ā and stops when it flattens.
š” Why it matters
EBRs hit 99.9% on Sudoku-Extreme with just 1.8M parameters ā using 345Ć fewer inference FLOPs than the previous best, EqR. On N-Queens, 20 samples cover 98% of all valid solutions; GRAM manages only 90%.
ā” Our take
Training recurrent reasoners at fixed depth is burning compute for nothing. An energy signal as a built-in progress meter is the smarter lever ā and finally answers the question of when a model is done thinking.
The title, summary and analysis of this item were produced automatically by an AI system and have not been editorially reviewed. They may contain errors, bias or omissions ā when in doubt, read the linked original source.


