Energy-Based Reasoners hit 99.9% on Sudoku-Extreme with 345× less inference compute

Researchers William Xingxu Chen, Alexi Gladstone, Zeyu Michael Li, Heng Ji and Yilun Du have introduced Energy-Based Reasoners (EBRs), a type of looped reasoning model. The work is cited as an arXiv preprint (arXiv:2611.08977) under the title "Energy-Based Reasoners: Your Recurrent Reasoner Should Be an Energy-Based Model".

The starting point is looping: a recurrent block is iterated repeatedly, so a model can scale its effective depth and reason over several steps in latent space. The authors say existing looped models are not guided by an explicit measure of progress. That makes it hard to know when to stop looping and how to improve the current state.

EBRs add that measure. They learn an energy function over latent reasoning states, and the looped updates are formulated as gradient descent on this energy. The energy value rates how compatible the reasoning state is with the initial board, and lower means better compatibility. It scores candidate predictions without an external verifier. Its gradient points in an improvement direction for refinement and also acts as a stopping signal: once the gradient is small, the state is stationary. Each trajectory stops when its energy gradient falls below a threshold, so easy puzzles finish in a few steps and hard ones keep going. When several candidate solutions exist, the one with the lowest energy is picked. Running more descent steps refines each answer further, and running more trajectories gives the energy more candidates to choose from.

Training uses a further ingredient, exploration-based training, which is meant to discover additional modes and to shape the energy landscape over a broader set of states.

The headline result is on Sudoku-Extreme. With only 1.8M parameters, EBRs reach 99.9% accuracy while using 345× less inference compute than the next-best model, and training compute stays comparable. Across four algorithmic reasoning benchmarks, the authors say EBRs match or outperform prior recurrent reasoning models.

On the multi-solution benchmarks, including N-Queens and Graph Coloring, EBRs reach state-of-the-art coverage of distinct valid solutions while keeping near-100% single-sample accuracy. Coverage is the share of each input's distinct valid solutions that 20 samples find. The authors describe EBR as acting like stochastic gradient descent that can capture many different solutions. Their conclusion is hedged: the results suggest that energy-based models are a promising approach for strong reasoning performance.

Key facts

  • Energy-Based Reasoners (EBRs) learn an energy function over latent reasoning states and treat looped updates as gradient descent on it; the gradient also decides when to stop.
  • With 1.8M parameters, EBRs report 99.9% accuracy on Sudoku-Extreme using 345× less inference compute than the next-best model.
  • Training compute is described as comparable to prior recurrent reasoning models, not reduced.
  • On N-Queens and Graph Coloring, EBRs report state-of-the-art coverage of distinct valid solutions (measured with 20 samples) with near-100% single-sample accuracy.
  • The authors are Chen, Gladstone, Li, Ji and Du; the work is cited as arXiv preprint 2611.08977.

Why it matters

Looped models gain depth by repeating one block, but they have lacked a built-in sense of progress. EBRs supply one: the energy scores a state, its gradient says how to improve it, and a small gradient says to stop. That lets easy puzzles finish in a few steps while hard ones keep going, which is where the reported inference savings on Sudoku-Extreme come from. The 345× figure is against the next-best model on that benchmark. The tiny model size, 1.8M parameters, makes the result notable.

Who it affects

Mainly researchers working on recurrent and looped reasoning models, and those exploring energy-based models as a route to stronger reasoning. The tested tasks are algorithmic puzzles: Sudoku-Extreme, N-Queens and Graph Coloring are named, out of four benchmarks in total.

How to use it

This is a research result, not a product. The page gives a method description and a citation, with the preprint at arXiv:2611.08977. Teams studying looped reasoners can take the design idea: learn an energy over latent states, run gradient descent on it, stop when the gradient is small, and pick the lowest-energy candidate among several samples. No release of code, weights or a demo is mentioned in the source text.

How solid is it

The claims come from the authors' own project page and are cited as an arXiv preprint. The source does not say the paper is peer-reviewed. The 99.9% accuracy and 345× figures are stated for Sudoku-Extreme. For N-Queens and Graph Coloring the page gives only "near-100%" and "state-of-the-art", with no exact scores. The next-best model is not named.

Risks and caveats

The 345× figure is a compute comparison on one benchmark against an unnamed next-best model. No absolute inference numbers (FLOPs, time, energy use) are given, and "compute" is not stated as electrical energy. Training compute is comparable, not lower. The fourth benchmark and the other baselines are not named. The authors' own conclusion is only that the results suggest energy-based models are a promising approach; the page does not claim results beyond these algorithmic reasoning tasks.

“These results suggest that energy-based models are a promising approach for strong reasoning performance.”

— Energy-Based Reasoners project page