1Duke University2UIUC3Harvard University
arXiv CodeTL;DR – We current Energy-Based Reasoners (EBRs), reasoning models that study an energy function complete latent reasoning states and formulate looped updates as gradient decline on this energy. Its value scores applicant predictions without requiring an external verifier, during its gradient guides refinement and provides a natural indication for stopping. With lone 1.8M parameters, EBRs accomplish 99.9% accuracy on Sudoku-Extreme during using 345× small conclusion compute than the next-best model. On N-Queens and Graph Coloring, EBRs accomplish state-of-the-art safety during maintaining near-100% single-sample accuracy.
Abstract
Looping is a mighty inductive bias for reasoning. By often iterating a recurrent block, looped models can measure their productive degree and allow multi-step reasoning in latent space. However, existing looped models are not inherently guided by an definitive measure of progress, making it difficult to decide whenever to halt looping and how to enhance upon the current state. To allow this, we current Energy-Based Reasoners (EBRs), which study an energy function complete latent reasoning states and formulate looped updates as gradient decline on this energy. The energy scores latent states, during its gradient guides refinement and provides a indication for stopping. To train EBRs effectively, we additional current exploration-based training to detect additional modes and to improved form the energy landscape complete a broader set of states. Across four algorithmic reasoning benchmarks, EBRs equivalent or outperform previous recurrent reasoning models, achieving 99.9% accuracy on Sudoku-Extreme during using 345× small conclusion compute than the next-best example and maintaining comparable training compute. On all multi-solution benchmarks, EBRs accomplish state-of-the-art safety of distinct valid solutions during maintaining near-100% single-sample accuracy. These results propose that energy-based models are a promising method for powerful reasoning performance.
Why Energy?
Existing looped reasoners update their latent province without any definitive appraisal of it. An EBR learns an energy function that guides refinement, stopping and selection.
Score. The energy rates the compatibility between the reasoning province and the first commission . Lower method improved compatibility.
Refine. The gradient turns the mark into an betterment direction.
Stop. The province becomes stationary formerly the gradient is small, so the gradient acts as a stopping criterion.
Select. Among multiple applicant solutions, we choice the one alongside the lowest energy.
Correct solutions resolve at low-energy stationary states. Energy, gradient norm and stage size during decline on Sudoku-Extreme. Correct trajectories resolve at low energy alongside small gradients, during incorrect ones keep moving.
Test-Time Scaling
Running additional decline steps refines all answer further, and operating additional trajectories gives the energy additional candidates to choose the lowest-energy one from.
EBRs attain EqR's finest accuracy alongside 7× small iterations (left) and 43× small samples (right).
GRAM and TRM from GRAM's paper; EqR from its released checkpoint. Mean ± SD complete three seeds.
Adaptive Stopping
Each trajectory stops formerly its energy gradient falls below a threshold, so uncomplicated puzzles complete in a few steps and difficult ones keep going.
| Method | # Params | Sudoku-Extreme ↑ | Maze-Hard ↑ |
|---|---|---|---|
| Single example (%) | |||
| TRM | 5–7M | 87.4 | 85.3 |
| EqR | 5–7M | 93.0 | – |
| FPRM | 7M | 94.2 | 87.0 |
| GRAM | 10M | 86.6 | – |
| Looped flows | 5–7M | 97.9 | 86.7 |
| EBROurs | 1.8M | 99.4 | 87.4 |
| Multiple samples (%), figure of samples N in parentheses | |||
| PTRM | 5–7M | 98.8(100) | 86.7(100) |
| EqR | 5–7M | 99.8(128) | – |
| GRAM | 10M | 96.1(20) | – |
| Looped flows | 5–7M | 99.3(5) | 86.9(5) |
| EBROurs | 1.8M | 99.9(5) | 88.3(5) |
EBRs provision the finest single-sample accuracy among reasoning models. Within lone 5 passes, they attain 99.9% and 88.3%, onward of methods that diagram up to 128. Stopping all trajectory formerly its energy gradient falls below a threshold keeps 99.9% accuracy alongside 10.9× small FLOPs than a fixed-depth EBR, allowing us to surpass EqR alongside 345× small FLOPs.
Multiple Solutions
EBR acts akin stochastic gradient decline and can grasp many distinct solutions. Coverage captures the portion of all input's distinct valid solutions that 20 samples find.
| Method | N-Queens | Graph Coloring | ||
|---|---|---|---|---|
| 8×8 | 10×10 | 8-vertex | 10-vertex | |
| Single sample | Pass@1 (%) ↑ | Conflicts ↓ | ||
| TRM | 66.8 | 17.5 | 109.3 | 170.7 |
| GRAM | 99.5 | 89.7 | 8.3 | 31.3 |
| EBROurs | 100 | 99.5 | 0 | 0 |
| 20 samples | Cov@20 (%) ↑ | Cov@20 (%) ↑ | ||
| TRM | 22.6 | 2.0 | 22.3 | 6.8 |
| GRAM | 89.7 | 57.5 | 85.8 | 51.3 |
| EBROurs | 98.1 | 92.0 | 88.6 | 54.5 |
Pass@1 (%) and Conflicts (edges summed complete the test graphs) use one sample. Cov@20 is the portion of distinct valid solutions established by 20 samples (%). Baselines from GRAM's paper.
Graph Coloring, 10 vertices
@misc{chen2026energybasedreasoners, heading = {Energy-Based Reasoners: Your Recurrent Reasoner Should Be an Energy-Based Model}, author = {Chen, William Xingxu and Gladstone, Alexi and Li, Zeyu Michael and Ji, Heng and Du, Yilun}, twelvemonth = {2026}, eprint = {2611.08977}, archivePrefix = {arXiv}, primaryClass = {cs.LG}, url = {https://arxiv.org/abs/2611.08977} }