the short version
- The headline efficiency number compares EBRs with one baseline, EqR, on one benchmark, Sudoku-Extreme. The body gives it as 345x, not the 385x in the page title.
- Accuracy holds at the speedup. A 1.8M-parameter EBR scores 99.9% on Sudoku-Extreme with 5 samples, against EqR's 99.8% with 128 samples.
- Stopping each trajectory once its energy gradient is small keeps 99.9% accuracy with 10.9x fewer FLOPs than a fixed-depth EBR. The page links this to the 345x result but does not break the figure down further.
- All four benchmarks are constrained algorithmic puzzles. The project page reports no wall-clock latency, hardware or absolute FLOP counts.
Energy-based reasoners (EBRs) are small looped models. They learn a scalar energy over latent reasoning states and run each loop step as gradient descent on that energy. The project page title says 385x, but the abstract and results section give the figure as 345x fewer FLOPs than EqR, the next-best model, on Sudoku-Extreme only. Accuracy holds at that speedup: 99.9% for EBRs against EqR's 99.8%.
The work comes from William Xingxu Chen, Alexi Gladstone, Zeyu Michael Li, Heng Ji and Yilun Du, and is listed as arXiv 2611.08977. It targets a specific weakness of looped reasoners. They iterate a recurrent block to scale effective depth, but nothing tells them whether the current state is improving or when to stop. EBRs supply that signal through a learned energy and are evaluated on four algorithmic benchmarks: Sudoku-Extreme, Maze-Hard, N-Queens and Graph Coloring.
What energy-based reasoners do at inference time
The model learns a function E_theta(z; x). It maps a latent reasoning state z and an input x, such as a Sudoku board, to a real number. Lower energy means the state is more compatible with the input. The project page describes four uses of that one function at inference time.
- Score: the energy rates how compatible the current reasoning state is with the initial board.
- Refine: the next state is z_{t+1} = Pi_M(z_t - alpha * grad_z E_theta(z_t; x)), a gradient step followed by a projection Pi_M.
- Stop: a trajectory halts when the gradient norm drops below a threshold tau, because the state has become stationary.
- Select: when several trajectories run, the answer is the candidate with the lowest energy.
The page reports that correct Sudoku-Extreme trajectories settle at low energy with small gradients, while wrong ones keep moving. That pattern is why the gradient norm can serve as a stopping rule. For training, the authors introduce what they call exploration-based training. Its stated aim is to discover additional modes and shape the energy landscape over a broader set of states. The project page gives no further detail on how it works.
Commenters on Hacker News asked what "energy" means here. One noted that the related work does not mention the free energy principle, and the sources do not mention it either. On the project page, the energy is a learned scalar in R that scores states, supplies a descent direction and signals when to stop. One commenter described EBRs as a pairwise weight matrix over Sudoku cells, searched by picking the lowest-energy discrete swap at each step. The project page describes something different: gradient steps on latent reasoning states followed by a projection.
The 345x figure compares against EqR on Sudoku-Extreme only
The efficiency claim has a single baseline and a single task. The abstract reports 99.9% accuracy on Sudoku-Extreme "while using 345× less inference compute than the next-best model." The results section names that model as EqR and gives the unit as FLOPs. The 385x in the page title does not appear anywhere in the body, so 345x is the figure with supporting text.
The page reports three related efficiency figures:
- Adaptive stopping keeps 99.9% accuracy with 10.9x fewer FLOPs than a fixed-depth EBR. Easy puzzles finish in a few steps and hard ones keep going. The page says this is what allows EBRs to surpass EqR with 345x fewer FLOPs.
- EBRs reach EqR's best accuracy with 7x fewer iterations.
- EBRs reach EqR's best accuracy with 43x fewer samples.
The page does not show how these factors combine into 345x. It does not report absolute FLOP counts, wall-clock latency, hardware or batch settings. It says training compute is "comparable" without giving a number. In the test-time scaling comparison, EqR figures come from EqR's released checkpoint, while GRAM and TRM figures come from GRAM's paper, so not every baseline was re-run under the same conditions. Capacity planning from these numbers would need the paper's FLOP accounting, which the project page does not include.
Accuracy holds at 1.8M parameters and five samples
With one sample, the EBR scores 99.4% on Sudoku-Extreme and 87.4% on Maze-Hard. The next-best single-sample results are Looped flows at 97.9% on Sudoku and FPRM at 87.0% on Maze. With five samples, the EBR reaches 99.9% and 88.3%. The closest multi-sample results are EqR at 99.8% with 128 samples on Sudoku and Looped flows at 86.9% with 5 samples on Maze. The EBR has 1.8M parameters, while the listed baselines range from 5M to 10M.
On Sudoku-Extreme, the faster method is also slightly more accurate, so the reported speedup comes with no accuracy loss. On Maze-Hard, the margin is 0.4 points with one sample and 1.4 points with multiple samples, and the page reports no efficiency comparison for that benchmark. The abstract says EBRs match or outperform prior recurrent reasoning models, which is consistent with the table.
Sampling covers more valid solutions than GRAM
The authors write that an EBR "acts like stochastic gradient descent and can capture many different solutions." They measure this with Cov@20, the share of each input's distinct valid solutions found by 20 samples. Graph coloring single-sample results are reported as conflicts, meaning edges summed over the test graphs. All baselines in this comparison come from GRAM's paper.
- N-Queens 8x8: Pass@1 of 100% (GRAM 99.5%), Cov@20 of 98.1% (GRAM 89.7%).
- N-Queens 10x10: Pass@1 of 99.5% (GRAM 89.7%), Cov@20 of 92.0% (GRAM 57.5%).
- Graph Coloring, 8 and 10 vertices: 0 conflicts in both cases (GRAM 8.3 and 31.3), Cov@20 of 88.6% and 54.5% (GRAM 85.8% and 51.3%).
The largest coverage gap is on 10x10 N-Queens. On 10-vertex graph coloring, EBRs and GRAM each find only about half the distinct solutions, down from 88.6% and 85.8% at 8 vertices. The page does not report compute for these multi-solution runs.
All four benchmarks are algorithmic puzzles, not language tasks
Commenters on Hacker News returned repeatedly to scope. One commenter noted that "reasoning" here means a sequence of intermediate passes in puzzle-solving models, not language models. Another wrote that the approach works "only on sudoku and basic games, with constrained rules, unique or low number of solutions" and asked how it would do on Chess or Go. The sources support the narrow reading. All four benchmarks are algorithmic puzzles, and the page makes no claim about language, open-ended tasks or larger state spaces.
Efficient inference is useful when it comes with something useful.
For engineers, the idea worth tracking is the control signal rather than the benchmark numbers. A single learned scalar ranks candidates, provides a descent direction and indicates when to stop. That combination is relevant wherever per-instance compute and best-of-N selection matter. Whether such an energy can be trained outside constrained puzzles is the open question, and the sources contain no evidence either way.
Questions this raises
What does the 345x compute saving for energy-based reasoners actually compare?
It compares inference FLOPs against EqR, the next-best model, on Sudoku-Extreme only. The project page does not show how the 10.9x, 7x and 43x factors combine into 345x. It also gives no absolute FLOP counts, wall-clock latency or hardware details.
How do energy-based reasoners decide when to stop iterating?
Each loop step is gradient descent on a learned scalar energy, followed by a projection. A trajectory halts when the gradient norm drops below a threshold tau. The page reports that correct Sudoku-Extreme trajectories settle at low energy with small gradients, while wrong ones keep moving.
Have energy-based reasoners been tested on language tasks?
No. All four benchmarks are algorithmic puzzles: Sudoku-Extreme, Maze-Hard, N-Queens and Graph Coloring. The page makes no claim about language, open-ended tasks or larger state spaces.
Spotted an error? Email samir.s.sengupta@gmail.com and I will correct it and date the change at the top of the note. How these notes are researched, drafted and checked is in the editorial policy.
