Energy-Based Reasoners
by Duke, UIUC and Harvard · Paper · Text & code
unproven
Tiny puzzle solver with big efficiency claims, untested outside narrow puzzles.
1.8M-parameter recurrent reasoner treating looped updates as energy minimization; page claims 345x less inference compute than next-best model on Sudoku-Extreme
StatusResearch
Why we say so
The authors claim a 1.8M-parameter model uses 345x less compute than the next best on Sudoku-Extreme, and have released code. HN commenters say it is limited to discrete puzzles and no substitute for LLM reasoning. One experienced commenter says energy-based models are hard to train and don't scale, and another noticed the headline says 385x while the page says 345x.
Evidence
- Nov 18A commenter with a year of experience says energy-based models are very hard to train, niche-specific, and sudoku solvers don't scale