Skip to main content
v2026.11,986 entries · CC-BY 4.0
CASRAIAsk CASRAI · AI assistant for researchAsk a research question. Answers come with their sources.Funder rules, compliance, methods and publishing. Answers link the sources they rest on.

Editorial · CASRAI · AI and ML research outputs

Preprint Shows Classical Energy Models Learn Reasoning Faster Than Looped Transformers

A November 2026 preprint shows classical energy models learn reasoning puzzles 1,000x faster than looped transformers with 282x fewer inference FLOPs.

Published 16 Nov 2026· 8 minute read

Ask CASRAI · free to try

Ask about this story

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

CASRAI is the reference for research administration — bookmark it for the next question.

A machine learning preprint released on November 12, 2026, demonstrates that classical energy-based neural architectures can learn algorithmic reasoning tasks over 1,000 times faster than modern looped transformer models while requiring orders of magnitude less training and inference compute. The study, authored by researchers at London-based AI lab Prefrontal, challenges prevailing assumptions regarding whether recurrent transformers are necessary for complex constraint-satisfaction problems across academic AI research.

What the preprint discovered: 1,000x training acceleration and 282x inference efficiency

The preprint, titled Classical Energies Learn to Reason 1,000x Faster than Looped Models (arXiv:2611.17309), was authored by Benjamin Hoover, Hugh Blayney, and James C. R. Whittington. The research team evaluated whether modern deep recurrent neural networks—frequently called looped models—are genuinely required to solve complex constraint-satisfaction puzzles such as Sudoku-Extreme and Maze-Unique.

Looped architectures have recently gained traction as a leading paradigm for iterative reasoning. Architectures such as the Hierarchical Reasoning Model (HRM) and Tiny Recursive Models (TRM) repeatedly apply a weight-tied neural layer to refine candidate answers. However, these systems require long training runs, suffer from optimization instability, and demand substantial GPU compute on relatively straightforward puzzles.

To establish a rigorous baseline, the Prefrontal team introduced CRAM (Compact Reasoning Associative Memory), a trainable variant of the continuous Hopfield and Tank energy formulation from 1985. The empirical outcomes demonstrated decisive advantages across core metrics:

  • Sudoku-Extreme: CRAM solved 99.996% of test puzzles after 42 milliseconds of training on a single NVIDIA H200 GPU, achieving 100% mean accuracy across test seeds.
  • Maze-Unique: CRAM solved 100% of test mazes after 132 milliseconds of training, outpacing looped transformer baselines by more than 1,000 times in training speed.
  • Inference compute: At matched accuracy, CRAM required at least 282 times fewer inference floating-point operations (FLOPs) than the strongest energy-based recurrent baseline, and up to 128,000 times fewer FLOPs than deep equilibrium models.
  • Parameter footprint: CRAM required only 7,700 non-zero parameters on Sudoku and 18 parameters on mazes, compared to 5 million to 27 million parameters in leading transformer models.

These findings demonstrate that for algorithmic tasks, rigorous measurement of a compute FLOPs estimate reveals massive computational redundancies in multi-layer deep neural networks.

Architectural mechanics: Hopfield-Tank energy formulation versus looped recurrence

The core innovation of CRAM lies in replacing unconstrained forward neural recurrence with an explicit mean-field free energy landscape. Reasoning tasks are structured over M answer cells, each assigned a probability distribution pi over C candidate values via a softmax activation with inverse temperature β applied to continuous neuron states vi.

The total energy of a candidate solution combines negative entropy, pairwise cell interactions, and external constraints:

Eθ(p) = β-1 ∑i pi⊤ log pi + 0.5 ∑i ≠ j pi⊤ Kij pj + ∑i Bi⊤ pi

Here, the matrix Kij contains learnable pairwise costs between candidate values, while bias vector Bi encodes clues and boundary conditions. When cell beliefs are fully determined, the entropy term vanishes, reducing energy to pairwise costs and clue agreements.

Inference proceeds via dynamical relaxation. Rather than executing a feedforward transformer stack, CRAM updates cell states along the negative gradient of the energy function:

τi (dvi / dt) = -vi – ∑j Wij pj – Bi ≃ -∇pi Eθ(p)

where Wij = (Kij + Kji⊤) / 2 represents symmetric interaction weights. In continuous time, energy is guaranteed never to increase. CRAM implements focused Gibbs search—a stochastic descent technique prioritizing cells with high local energy—allowing the model to escape shallow local minima.

Crucially, this explicit energy formulation provides autonomous early stopping. While looped transformers generally run a fixed iteration budget or require an auxiliary halting network, CRAM halts inference immediately when board energy falls below an empirical threshold fitted during training, or when single-cell updates yield no further energy reduction.

Comparative performance across algorithmic reasoning benchmarks

The Prefrontal team benchmarked CRAM against leading recurrent and looped architectures on identical hardware:

Model Architecture Sudoku Parameters Sudoku Solved (%) Sudoku Fit Time Sudoku FLOPs / Puzzle Maze Solved (%) Maze Fit Time
HRM (Wang et al., 2026) 27.0M 54.62% ± 4.18% 1.2 hours 2.20 × 1011 39.97% ± 33.90% 7.1 hours
TRM (Jolicoeur-Martineau, 2025) 5.0M 85.73% ± 0.74% 2.5 hours 4.11 × 1011 92.23% ± 13.37% 15.6 hours
GRAM (Baek et al., 2026) 9.1M 96.10% 3.2 hours 8.51 × 1012 98.63% ± 1.31% 23.2 hours
EqR (Huang et al., 2026) 5.0M 99.49% ± 0.14% 2.0 hours 2.11 × 1014 93.00% 7.2 hours
FRM (Freinschlag et al., 2026) 7.0M 99.13% ± 0.29% 2.9 hours 3.91 × 1012 97.27% ± 2.80% 9.7 hours
EBR (Chen et al., Nov 2026) 1.8M 99.85% ± 0.05% 2.8 hours 3.37 × 1011 89.80% ± 6.35% 2.0 hours
CRAM (Learned Energy) 7.7K 100.00% ± 0.00% 0.04 seconds 2.02 × 109 100.00% ± 0.00% 0.13 seconds
CRAM (Hand-Set Energy) 1 unique 100.00% ± 0.00% None (0.00 s) 1.08 × 108 100.00% ± 0.00% None (0.00 s)

CRAM was also evaluated on N-Queens and Graph Coloring, matching or surpassing neural baselines while fitting in fractions of a second. Standardized evaluation methodologies detailed in CASRAI guidance on performance profiles emphasize evaluating solver runtime distributions alongside raw accuracy.

Training dynamics: Local objectives bypass backpropagation through time

The primary reason looped neural networks require hours to train is their reliance on Backpropagation Through Time (BPTT). Calculating gradients across unfolded reasoning trajectories creates severe memory overhead and gradient instability.

CRAM sidesteps trajectory differentiation entirely by using local statistical objectives, primarily pseudolikelihood (PL) and minimum probability flow. Under pseudolikelihood, the training objective isolates individual cells, maximizing the conditional likelihood of the ground-truth cell value given observed puzzle solutions:

∇θ J(θ) = Ex ~ Pdata [∇θ Eθ(x)] – Ex– ~ Pθ [∇θ Eθ(x–)]

By fitting the energy landscape directly to valid puzzle states without simulating search paths, training completes in milliseconds. CRAM also demonstrates high data efficiency: on Sudoku-Extreme, it achieved 98% accuracy using an order of magnitude fewer unique training puzzles than looped models.

Furthermore, CRAM provides clear weight interpretability. Because it optimizes explicit pairwise interaction costs, the learned weights directly reflect problem constraints. On Sudoku, inspection of the learned Kij matrix revealed large positive energy penalties for matching digits in the same row, column, or 3×3 box, learned without explicit rule programming.

A critical caution: Benchmark saturation and the AI reasoning scaling debate

Despite CRAM’s impressive speedups, the authors issued an explicit caution regarding benchmark interpretation. High scores on Sudoku-Extreme and Maze-Unique have frequently been cited as proof that looped transformers are developing general algorithmic reasoning capabilities.

The Prefrontal researchers caution that constraint satisfaction puzzles like Sudoku are fundamentally pairwise graphical models that possess exact closed-form energy representations. The fact that a 7,700-parameter classical energy model solves Sudoku after 42 milliseconds of training demonstrates that these benchmarks do not measure hierarchical abstraction or general problem-solving. Relying on such a synthetic benchmark to claim architectural scaling breakthroughs is fundamentally misleading.

This critique aligns with independent evaluations. In August 2025, the ARC Prize Foundation analyzed recurrent reasoning architectures in their report on HRM performance drivers, noting that recurrent networks often exploit localized heuristics rather than mastering open-ended abstract reasoning. Similarly, concurrent research on Energy-based reasoners (Chen et al., arXiv:2611.08977) emphasized that while energy formulations stabilize reasoning trajectories, scaling beyond toy benchmarks requires learning representations in continuous latent spaces rather than discrete board representations.

What it means for academic computing, research offices, and AI evaluators

The findings presented in the preprint carry direct practical implications for university computing facilities and research administrators:

  • Compute resource allocation: University research offices and computing centers manage limited GPU budgets. Demonstrating that classical energy formulations solve constraint tasks in milliseconds indicates that significant academic compute allocations may be misdirected toward brute-force transformer loops.
  • Evaluation standards in peer review: Program committees for major machine learning conferences must enforce stricter baseline requirements. Studies asserting reasoning breakthroughs on Sudoku, mazes, or grid pathfinding should be required to compare against classical energy and constraint acquisition baselines before publication.
  • Open science and reproducibility: The research team published their full codebase and configuration scripts in an open repository via Prefrontal AI on GitHub. In accordance with best practices for open code, the minimal computational footprint of CRAM allows independent academic laboratories to replicate findings on standard workstation hardware without requiring institutional cluster grants.

What happens next: Latent energies and hybrid reasoning architectures

The Prefrontal researchers outlined two primary next steps for energy-based reasoning. First, to overcome the limitations of pairwise formulations on non-local dependencies (such as complex multi-branch maze paths or mathematical proofs), the authors are extending CRAM with latent variables and hierarchical energy layers.

Second, researchers are investigating hybrid architectures where foundation models generate candidate reasoning sketches and lightweight energy modules perform deterministic constraint verification and error correction at inference time. As AI researchers continue to evaluate test-time compute, the CRAM preprint demonstrates that classical neural network principles offer powerful alternatives to unconstrained recurrence.

Sources

  • Hoover, B., Blayney, H., & Whittington, J. C. R. (2026). Classical Energies Learn to Reason 1,000x Faster than Looped Models. arXiv preprint arXiv:2611.17309. Checked 2026-11-16. https://arxiv.org/abs/2611.17309
  • Prefrontal AI. (2026). CRAM: Compact Reasoning Associative Memory (Source Code Repository). GitHub. Checked 2026-11-16. https://github.com/Prefrontal-AI/CRAM
  • Chen, W. X., Gladstone, A., Li, Z. M., Ji, H., & Du, Y. (2026). Energy-based reasoners: Your recurrent reasoner should be an energy-based model. arXiv preprint arXiv:2611.08977. Checked 2026-11-16. https://arxiv.org/abs/2611.08977
  • ARC Prize Team. (2025). The hidden drivers of HRM’s performance on ARC-AGI. ARC Prize Foundation. Checked 2026-11-16. https://arcprize.org/blog/hrm-analysis
  • Jolicoeur-Martineau, A. (2025). Less is more: Recursive reasoning with tiny networks. arXiv preprint arXiv:2510.04871. Checked 2026-11-16. https://arxiv.org/abs/2510.04871

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →