345x more efficient, but only at Sudoku: energy-based reasoners explained
Two new papers report that tiny energy-based models solve puzzles with hundreds of times less compute than looped models. We explain how "thinking" turns into rolling downhill through a landscape, and we checked the 345x math ourselves. The second paper also suggests the Sudoku benchmark itself may be the problem.

What's going on
Two papers came out in the same November week that sound almost identical: energy-based models supposedly need hundreds of times less compute to solve puzzles than so-called looped models. The first comes from William Chen, Alexi Gladstone, Zeyu Li, Heng Ji and Yilun Du (Duke, UIUC, Harvard) and is called Energy-Based Reasoners (EBR). It cites 345 times less compute. The second is from Benjamin Hoover, Hugh Blayney and James Whittington at the London lab Prefrontal and is called CRAM. Its title alone promises "1,000x faster." On Hacker News, where the EBR post ranked high with 283 points, the headline even said "385x." One commenter pointed out that the project page says 345.
Our thesis: the idea behind both papers is elegant and easy to follow. But the efficiency headlines only hold within a very small league: tiny models solving Sudokus and mazes of fixed size. And the second paper itself makes the strongest case for just how small that league is.
What a learned "energy" is
Our explainer on looped models covered networks that run the same block of layers multiple times, like a tool you apply to a workpiece again and again. The mini reasoners TRM and HRM work exactly this way: they refine an internal solution draft loop after loop. According to the EBR paper, though, these models lack an explicit measure of how far along they are. So it's unclear when they should stop.
EBR teaches the model an energy function. That's a number indicating how well a solution draft fits the given puzzle. Low energy means "fits"; high energy means "breaks the rules." Picture a hilly landscape: every point is a possible intermediate state, the height is the energy, and good solutions sit in the valleys. Thinking then means rolling downhill. Technically, each loop is a step of gradient descent, where you always move in the direction of steepest descent.
According to the project page, the energy does four jobs:
- Score: It tells you how good the current intermediate state is.
- Improve: Its slope shows which direction the next step should go.
- Stop: When the terrain flattens out, meaning the slope drops below a threshold τ, the model is done. That way it notices on its own that it has arrived.
- Select: Across multiple solution attempts, the one with the lowest energy wins. No external verifier needed.
On top of that comes a training trick called Explorative Modeling. The model rolls down several paths in parallel so the landscape is shaped sensibly even away from the first solution it finds. That helps with puzzles that have many valid solutions, like N-Queens. Under the hood, EBR consists of just two transformer blocks with 1.8M parameters in total. It was trained on a single Nvidia H200, roughly 3 to 4 hours for Sudoku according to the README. The code is on GitHub under the MIT license, with checkpoints on Hugging Face.
The authors also spell out the costs: the descent requires second-order derivatives, which are more expensive computations than a directly learned transition. Plus there are extra knobs to tune, like step size and noise level.
Where the 345 comes from: we did the math
The project page has no absolute numbers, but the paper's appendix does. The 345x applies only to inference, meaning solving, not training. The comparison is against EqR (Equilibrium Reasoners, Kolter group, ICML 2026). EqR only reaches its 99.8% on Sudoku-Extreme with 128 attempts of 64 steps each, which according to the EBR paper costs about 210 TFLOPs per puzzle (FLOPs count arithmetic operations). EBR gets by with 8 attempts and adaptive stopping at 0.61 TFLOPs. 210 divided by 0.61 is 344, so the 345x checks out.
| Setting (Sudoku-Extreme) | Avg. steps | Accuracy | TFLOPs per puzzle |
|---|---|---|---|
| EqR, 128 attempts × 64 steps | 64 | 99.8% | ≈ 210 |
| EBR, 8 attempts, no stopping | 64 | 99.92% (Pass@8) | 6.60 |
| EBR, 8 attempts, stopping τ = 0.2 | 5.6 | 99.92% (Pass@8) | 0.61 |
| EBR, 1 attempt, stopping τ = 0.2 | – | 99.4% | 0.060 |
Two things stand out. First, much of the lead comes from EqR launching a massive number of attempts to hit its best scores. Second, self-stopping alone accounts for a factor of 10.8 (the project page says 10.9). Easy puzzles finish after a few steps; hard ones run longer. With a coarser threshold, the savings climb to 17 to 30 times, but accuracy then collapses to 74% at τ = 5.
The "comparable training budget" from the abstract mainly holds for Sudoku: there, according to its own table, EBR needs 1.92 instead of 2.26 EFLOPs compared with the baseline GRAM. For N-Queens and graph coloring, however, it takes 1.2 to 3.7 times more training compute than GRAM.
The second paper: 7,700 parameters are enough
CRAM goes the opposite way. Instead of learning an energy over a transformer's internal states, it falls back on an energy that John Hopfield and David Tank designed for logic puzzles back in the 1980s. All it learns are pairwise "costs" between cells: how bad is it if cell A and cell B hold the same digit? Training is mostly done via pseudolikelihood, without backpropagation through time, meaning without the expensive backward pass through every loop.
The results, according to the paper:
- Sudoku-Extreme: 99.996% solved, with about 7,700 nonzero parameters, after 42 milliseconds of training on an H200.
- Maze-Unique: 100% with 18 parameters, after 132 ms of training.
- Maze-Hard: 99.9% with a 29-parameter extension, although its hidden layer was chosen using knowledge of the task.
So the "1,000x faster" in the title refers to training time: milliseconds versus hours; TRM, for example, takes 2.5 hours. The authors concede that their times were measured after warm-up, while the baselines' times came from cold starts, on a shared cluster. For inference, CRAM reports at least 282 times fewer FLOPs than EBR, the cheapest baseline at top accuracy. For Maze-Hard, the CRAM table lists 1.41 × 10⁷ FLOPs per maze versus 1.93 × 10¹³ for EBR. Oddly, according to this locally computed table, EBR there even needs five times more than TRM (3.84 × 10¹²).
An independent reimplementation on GitHub (catid/hopfield) confirms the accuracy figures. An energy that merely counts rule violations solves 99.96% of the Sudokus there without any optimizer at all. But the reimplementation couldn't independently confirm the 1,000x speed claim on its own Blackwell hardware. The paper's real takeaway is a warning: high scores on these logic benchmarks don't show that a model scales to harder tasks. Its own simple energy would fail on ARC-AGI.
The Sudoku league and the missing yardstick
Sudoku-Extreme and Maze-Hard come from Sapient's HRM paper (2025). TRM from Samsung SAIL with 7M parameters followed, along with EqR, GRAM and others. The ARC Prize's review of HRM already found that the hierarchical architecture contributed little. A regular transformer of the same size trailed by only about 5 percentage points; most of the gain came from the outer refinement loop and data augmentation. Sudoku-Extreme consists of 1,000 training puzzles, each transformed 1,000 times via symmetries.
In this league, everyone only competes with each other. EBR never goes up against a language model or a classic solver. According to its own benchmark table, the open-source solver tdoku handles the roughly 49,000 hardest forum Sudokus at 41.7 microseconds per puzzle on an Ice Lake laptop. Collections like these are exactly what Sudoku-Extreme draws on. Our rough extrapolation: the 422,786 test puzzles would take about 18 seconds on a single laptop core. For EqR, the EBR paper puts the same test run at about 120 GPU hours on an H200. That's not a fair contest, since tdoku knows the rules while the networks learn them from examples. But it shows how small the numbers inside the league look compared with the outside world. On Hacker News, user program_whiz estimated a classic CPU approach would be "probably 10 to 100 times faster" and placed the method in the "deterministic, fixed problem size" category. User mocmoc wrote that such models are "insanely hard to train."
On top of that, every model is trained for exactly one size: 9×9 Sudoku, 8×8 and 10×10 Queens, graphs with 8 and 10 nodes. Whether a model that learned 10×10 can also handle 12×12 isn't tested in either paper.
Kona and the question of language models
The name Logical Intelligence came up quickly in the HN discussion. The company unveiled its energy-based model Kona 1.0 in January 2026; Yann LeCun chairs its Technical Research Advisory Board. There's no link to the EBR paper: the paper doesn't mention the company, and the acknowledgments list no funding from it. According to its own blog, Kona solved 96.2% of "hard" Sudokus in a public demo at an average of 313 ms. GPT-5.2, Claude Opus 4.5 and other frontier models managed about 2% combined. But the LLMs weren't allowed to execute code, meaning they couldn't write a solver. The page doesn't say where the puzzles came from, nor how big the model is. According to an interview with ByteByteGo, Kona has "ten to a few hundred million" parameters. We found no independent tests, no paper and no published results from the announced pilot projects with industry partners. There are no open weights and no API.
And language models? There are two bridges, both still narrow. Blondel et al. show mathematically in "Autoregressive Language Models are Secretly Energy-Based Models" (ICML 2026) that ordinary token-by-token models and energy-based models can be converted into one another. That's a theoretical equivalence, not a faster method. More practical are the Energy-Based Transformers by Gladstone et al. (ICLR 2026 Oral), co-authored by EBR's Alexi Gladstone. According to the authors, they scale up to 35% better during training than a standard transformer. But they were only tested up to 800M parameters, and the authors themselves write that training and inference are more expensive than for regular models and that poorly chosen settings often lead to unstable training.
The bigger picture
For researchers, it's been a good week: EBR turns the looped model into something that can measure its own progress and stop on its own. Compared with the core problem from our looped-model explainer, that every loop costs compute, that's real progress. In our view, the 10x savings from self-stopping is the more robust result. The 345x leans heavily on the baseline needing 128 attempts for its best score.
CRAM puts the headlines in perspective. If 7,700 parameters and 42 milliseconds of training are enough for 99.996%, Sudoku-Extreme mostly measures whether a model can pick up simple pairwise rules, not whether it "thinks." Our take: none of these efficiency numbers transfer to language models, agents or open-ended tasks. They apply to fixed, fully rule-based puzzles for which faster classic solvers have existed for decades. Things only get exciting once energy-based reasoners solve tasks where nobody can write down the rules in advance, such as ARC-AGI. Both papers skipped that test, and according to its README, EBR will release the ARC checkpoints only "soon."
Sources
- Projektseite Energy-Based Reasoners energy-based-reasoners.github.io
- arXiv: Energy-Based Reasoners (2611.08977) arxiv.org
- arXiv-HTML: Energy-Based Reasoners arxiv.org
- GitHub: wwillchen/ebr github.com
- arXiv: Classical Energies Learn to Reason 1,000x Faster than Looped Models (2611.17309) arxiv.org
- arXiv-HTML: CRAM-Paper arxiv.org
- Prefrontal: Reasoning at Hyperspeed prefrontal.ai
- GitHub: Prefrontal-AI/CRAM github.com
- GitHub: catid/hopfield (unabhängiger Nachbau) github.com
- GitHub: t-dillon/tdoku github.com
- ARC Prize: The Hidden Drivers of HRM's Performance on ARC-AGI arcprize.org
- arXiv: Less is More – Recursive Reasoning with Tiny Networks (TRM) arxiv.org
- arXiv: Equilibrium Reasoners (EqR) arxiv.org
- Logical Intelligence: Kona-Sudoku-Demo logicalintelligence.com
- ByteByteGo: How to Build Models Without Language, with Logical Intelligence blog.bytebytego.com
- arXiv: Energy-Based Transformers are Scalable Learners and Thinkers arxiv.org
- arXiv: Autoregressive Language Models are Secretly Energy-Based Models arxiv.org
- CASRAI: Energy Models Learn Reasoning 1,000x Faster casrai.org
- kenashe.ai: CRAM makes Sudoku reasoning look too easy kenashe.ai
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.