This Tiny 7,700-Parameter Model Learns Reasoning 1,000x Faster Than Looped Transformers — and Questions the Benchmarks

Photo: Hippopx (public domain). Illustrative image.
What happened. A paper submitted to arXiv on 12 November 2026 — “Classical Energies Learn to Reason 1,000x Faster than Looped Models”, by Benjamin Hoover, Hugh Blayney and James C. R. Whittington of Prefrontal — revives a 1980s idea and reports startling numbers. CRAM, short for Compact Reasoning Associative Memory, is a learned version of the energy formulation Hopfield and Tank proposed in the 1980s. With just 7,700 parameters, it solved 99.996% of Sudoku-Extreme test puzzles after 42 milliseconds of training on a single GPU, and 100% of Maze-Unique puzzles after 132 milliseconds — at least 1,000 times faster to train than looped transformer models on these tasks. At matched accuracy, it used at least 282 times fewer inference FLOPs than the strongest energy-based recurrent baseline.
How it works. Instead of a transformer that loops over a problem step by step, CRAM treats reasoning as energy minimisation: the model learns an energy function over possible answers, then settles into the lowest-energy configuration — the solution. It is the classical Hopfield-network idea, rebuilt with modern learning. The tiny parameter count is the point: the structure of the problem is baked into the energy function, so there is very little to learn. Training takes milliseconds because the model is not learning reasoning from scratch; it is learning to use a representation that already encodes the puzzle’s constraints.
The caveat is the story. The paper’s most important passage is its own warning. Sudoku and Maze, the authors write, are “trivially solvable” pairwise graphical models with exact closed-form energy representations — which means they are almost perfectly suited to this approach by construction. The paper explicitly cautions against using these benchmarks to claim general reasoning ability. The 1,000x figure applies to these two puzzle benchmarks only, not to reasoning in general, and not to the messy, open-ended problems where today’s large models earn their keep. A model that solves Sudoku in 42 milliseconds has not demonstrated that it can plan, write code, or do science.
Why it still matters. Even with the caveat, CRAM is a useful provocation. It shows that for problems with the right structure, a tiny, specialised architecture can beat a giant general one by orders of magnitude on training time and compute — a reminder that scale is not the only lever in AI. The question it raises is a research question, not a product claim: which real problems have the kind of structure that an energy-based approach can exploit, and which do not? The same week, MIT researchers showed a model reading and editing other models’ raw weights, and Google’s AlphaProof Code formally verified the Signal protocol’s encryption — three very different answers to the question of what reasoning in AI actually is. For the background on how large models learn, see What Are Large Language Models? and What Is Fine-Tuning?.
Sources
- Hoover, Blayney & Whittington (Prefrontal) — “Classical Energies Learn to Reason 1,000x Faster than Looped Models”, arXiv:2611.17309 (submitted 12 November 2026): https://arxiv.org/abs/2611.17309
- FreshSynapse — “MIT’s Weight-Space Programming: GPT-6 Astra Reads and Edits Other Models’ Brains”: https://freshsynapse.com/articles/mit-weight-space-programming-gpt-6-astra-2026-11-15/
- FreshSynapse — “AI Proved Signal’s Encryption Secure in 4 Weeks — Work That Used to Take Years”: https://freshsynapse.com/articles/alphaproof-code-verifies-signal-encryption-2026-11-16/
- FreshSynapse — “What Are Large Language Models? A Plain-English Guide”: https://freshsynapse.com/articles/what-are-large-language-models/