AIToday
Large Language ModelsQiita 機械学習Published: Nov 19, 2026, 13:00 JST

EBR hits 99.9% on Sudoku-Extreme at 1.8M params

Send on LINE
EBR hits 99.9% on Sudoku-Extreme at 1.8M params

3 Key Points

  1. What happened

    Energy-Based Reasoners (EBR) report 99.9% accuracy on Sudoku-Extreme, reached with 1.8M parameters, while using 345× less inference compute than the next-best model; single-pass accuracy is 99.4%.

  2. Why it matters

    Because EBR uses the same energy score to update, stop, and pick among candidates, easy puzzles end in a few steps while only hard ones keep iterating — a design that may suit scheduling and constraint optimization where judging candidates is hard.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

EBR sits on a simple idea: instead of running a fixed number of update blocks and then applying a separate stopper or verifier, treat the model's state as a point on a landscape scored by an energy value and move it downhill. The paper's framing recasts iterative reasoning as energy minimization, so the same score that guides each update also tells the model when the trajectory has settled and which of several final candidates to take.

That contrasts with the prior chain of work it is placed against. Tiny Recursive Model and Hierarchical Reasoning Model also iterate to solve Sudoku and Maze, but they place stopping and selection outside the loop. EBR's claimed novelty is that update and scoring live in the same coordinate system, which is offered as the reason candidate comparison becomes principled rather than bolted on.

The reported numbers — 99.9% on Sudoku-Extreme, 1.8M parameters, and 345× less inference compute than the next-best model — are striking, but what is not yet visible: where the compute savings actually come from, whether a single gradient threshold holds across problems, and what the energy landscape looks like when a puzzle has multiple valid solutions. Those open questions, rather than the headline accuracy alone, are likely to decide how far the approach travels beyond Sudoku.

FAQ
How is EBR different from earlier iterative reasoning models?
Earlier models such as Tiny Recursive Model and Hierarchical Reasoning Model also solve Sudoku and Maze by iterating, but EBR defines the loop itself as energy minimization, so updating and scoring share one coordinate system.
Does EBR always use the same number of steps?
No. Each trajectory stops when its gradient falls below a threshold, so simple puzzles finish in a few steps while harder ones keep iterating.
What accuracy does EBR get in a single pass?
The 99.9% figure is for picking the lowest-energy answer out of five runs; a single run gives 99.4%.
Qiita 機械学習Read Original Article
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleEhime pearl farms turn to underwater sensors and AI