hn.today

Classical Energies Learn to Reason 1k× Faster Than Looped Models

arxiv.org20 points0 comments
Screenshot of Classical Energies Learn to Reason 1k× Faster Than Looped Models

Looped neural models solve reasoning tasks by repeatedly applying computation, but they can be unstable, expensive to train, and inefficient on simple puzzles. This work proposes CRAM, a learned energy-based system built on the classical Hopfield-Tank approach. Rather than training through a solver’s repeated steps, CRAM learns an energy function encoding puzzle constraints, then finds solutions by descending that energy. The approach supports early stopping and allocates computation according to problem difficulty, while its learned weights can be interpreted as representations of the constraints.

On test sets, CRAM solved 99.996% of Sudoku-Extreme puzzles and every Maze-Unique puzzle. The reported training times were 42 milliseconds for Sudoku and 132 milliseconds for mazes on one GPU - at least 1,000 times faster than the looped-model comparisons. For Sudoku-Extreme, it also matched looped models’ best accuracy while using at least 282 times fewer inference FLOPs. The authors argue that energy-based learning is a strong alternative for learning reasoning from data, combining speed with adaptive computation. They also caution that Sudoku and maze benchmarks may be too easy to establish whether looped reasoning methods scale: strong results on such tasks can reflect efficient constraint solving rather than the ability to handle harder forms of reasoning.

Read on arxiv.org0 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in AI

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.