CRAM learns to reason 1,000x faster than looped models — Learned energy descent — What does it mean?
The news. On November 12, 2026, researchers at Prefrontal posted Classical Energies Learn to Reason 1,000x Faster than Looped Models to arXiv. Their model, CRAM (Compact Reasoning Associative Memory), is a learned version of Hopfield and Tank's 1980s energy. Per the paper, it solves 99.996% of Sudoku-Extreme and 100% of Maze-Unique test puzzles after 42 ms and 132 ms of training on one H200 GPU, and matches looped baselines at their highest Sudoku-Extreme accuracy using at least 282× fewer inference FLOPs. The authors also warn that these puzzles are too easy to show that looped reasoning scales. Read the paper →
Picture the marble on the terrain. Every possible way to fill a Sudoku grid is a spot on the landscape, and the height of that spot is its energy. CRAM's whole job at inference time is to let the marble roll downhill until it lands in a valley that is low enough to count as solved. The terrain itself is simple: for every pair of cells, a learned small table of costs says how much height to add when cell i holds value a and cell j holds value b. Two equal digits in the same row add a bump. Add up all the pair costs plus a per-cell bias and you have the energy of the board.
Rolling downhill is concrete arithmetic. Each cell keeps a score per candidate value, the scores go through a softmax to become beliefs, and one update sets a cell's scores from the summed pair costs of its neighbours. One update is one layer of recurrence, exactly what a looped model does — but here each step follows the slope of a known energy, so in the noise-free version the energy never goes up. Because pure downhill can stall in a local dip, CRAM shakes the marble: it runs a noisy, Gibbs-style search that picks which cell to update by its local energy, and on Sudoku it releases 4 independent marbles from random starting boards.
The terrain also tells the marble when to stop. After training, CRAM measures the energy of every training solution and draws the stop line at the highest one plus a small tolerance. A board below that line is accepted, so an easy puzzle stops after a few steps and a hard one keeps rolling. Per the paper, this test never stopped on a wrong board in its experiments, while the separate halting rules of the looped baselines stopped on a wrong board on 0.05–3.54% of the puzzles they stopped on.
Training is where the contrast with looped models is sharpest. A looped model such as TRM reapplies a small Transformer-style block many times and learns by backpropagating through those repeats, then bolts on a halting head to decide when to stop. CRAM instead shapes the terrain directly: push the height down at each known solution and up at competing answers, without ever unrolling its own search. With the pseudolikelihood objective each training example is scored at the solution itself, which is why the paper reports 48 ms to reach 99% on a development set, while training the same energy with full BPTT reached only 48.8% of test puzzles after about 5 hours.
A side effect is that the weights are readable. Nobody told CRAM the rules of Sudoku, yet in every training run it put positive equal-digit costs on exactly the 810 cell pairs that share a row, column or box, and zero elsewhere. On mazes its 18 learned costs amount to dead-end filling, a classic maze algorithm. A looped model with millions of weights offers no such view.
| Model (Sudoku-Extreme) | Solved | Params | FLOPs per puzzle | Fit time | Source |
|---|---|---|---|---|---|
| HRM (looped) | 54.62% | 27.0M | 2.20×1011 | 1.2 h | Table 2 |
| TRM (looped) | 85.73% | 5.0M | 4.11×1011 | 2.5 h | Table 2 |
| EBR (energy on a looped transformer) | 99.85% | 1.8M | 3.37×1011 | 2.8 h | Table 2 |
| TRMP (looped) | 99.87% | 32.8K | 4.57×1012 | 9.0 h | Table 2 |
| CRAM (learned) | 99.996% (Table 2 rounds to 100.00%) | 7.7K | 2.02×109 | 0.04 s | Table 2 |
| CRAM (weights set by hand) | 100% | 1 unique weight | 1.08×108 | none | Table 2 |
Here is what that table means for one model pair, holding the task (the 10,000 Sudoku-Extreme test puzzles) and the hardware (one H200) fixed, at each model's own listed inference setting. TRM spends 4.11×1011 FLOPs per puzzle and solves 85.73%. CRAM spends 2.02×109, which is 4.11×1011 ÷ 2.02×109 ≈ 203× less compute per puzzle, while solving 99.996% versus TRM's 85.73%, with 7.7K nonzero parameters against 5.0M. This is not a matched-accuracy comparison; the paper's matched-accuracy figure is at least 282× fewer FLOPs. The FLOPs are also not spread evenly: CRAM's median compute grows 22× from the easiest difficulty bin to the hardest, while the median of the energy-on-transformer baseline EBR stays between 3.3 and 4.2×1011 in every bin. That is the spend more on harder problems idea, produced by the stop line rather than by a trained budget predictor. (Fit times come from a shared cluster, so the paper notes they vary with load.)
The most pointed result is the last table row. A CRAM energy whose few weights were set by hand from the puzzle rules solves every test puzzle on all four benchmarks (Sudoku, mazes, N-Queens and graph coloring), so a high score on Sudoku or mazes only shows that a model learned a handful of pairwise rules. The authors say plainly that this simple pairwise energy would struggle on harder tasks such as ARC-AGI. The lasting idea is not that pairwise energies replace looped transformers; it is that convergence, early stopping and cheap training come for free once the loop descends an explicit energy, instead of being added one module at a time.
Goes deeper in: AI Agents → Planning & Reflection → When to Spend More Tokens
Related explainers
- SMELT — looped depth reuse — how a large MoE model gets depth by running its middle layers twice
- UpLoop — upcycling into a looped model — training looped Transformers at the dense model's FLOPs
- ARLoop — attention over recurrent depth — letting each loop step look back at earlier depth states