基于能量的推理器:推理效率提升385倍
Energy-Based Reasoners: 385x more efficient inference

原始链接: https://energy-based-reasoners.github.io/

基于能量的推理器(EBRs)是循环推理模型,它学习潜在推理状态上的能量函数,并将循环更新作为对该能量的梯度下降。能量对状态兼容性进行评分,其梯度引导精炼并提供停止信号,低能量平稳态对应正确解。训练利用探索来塑造更多状态上的能量景观。在算法推理基准上,EBRs尽管仅使用1.8M参数,也能匹配或超越先前的循环模型:它们在Sudoku-Extreme上达到99.9%准确率,推理计算量比次优模型少345倍,并在N-Queens和图着色上保持接近100%的单样本准确率与最先进的解覆盖。基于梯度阈值的自适应停止进一步在保持准确率的同时降低计算。总体而言,基于能量的潜在推理提供了强大的性能和高效的测试时扩展。

Hacker News 讨论一篇论文,声称基于能量的推理器(Energy-Based Reasoners)推理效率提高385倍。评论者将EBM解释为在离散状态空间上进行迭代能量最小化,适用于数独、迷宫和图着色等谜题,其中冲突引导更新。然而,许多人指出其局限:问题必须有固定规则和尺寸,训练需要带标签样本,对模糊语言、图像或开放性推理的泛化尚不明确。一些人将EBM与LLM对比,LLM在语言和代码上更强,但在精确算术和计数上更弱。其他人争论基准测试选择,如禁用代码执行,并质疑可扩展性及相关公司的可信度。总体而言,该讨论认为离散推理有高效潜力,但怀疑其会广泛取代LLM。
相关文章

原文
Energy-Based Reasoners: Your Recurrent Reasoner Should Be an Energy-Based Model

1Duke University2UIUC3Harvard University

TL;DR – We introduce Energy-Based Reasoners (EBRs), reasoning models that learn an energy function over latent reasoning states and formulate looped updates as gradient descent on this energy. Its value scores candidate predictions without requiring an external verifier, while its gradient guides refinement and provides a natural signal for stopping. With only 1.8M parameters, EBRs achieve 99.9% accuracy on Sudoku-Extreme while using 345× less inference compute than the next-best model. On N-Queens and Graph Coloring, EBRs achieve state-of-the-art coverage while maintaining near-100% single-sample accuracy.

Abstract

Looping is a powerful inductive bias for reasoning. By repeatedly iterating a recurrent block, looped models can scale their effective depth and enable multi-step reasoning in latent space. However, existing looped models are not inherently guided by an explicit measure of progress, making it difficult to determine when to stop looping and how to improve upon the current state. To enable this, we introduce Energy-Based Reasoners (EBRs), which learn an energy function over latent reasoning states and formulate looped updates as gradient descent on this energy. The energy scores latent states, while its gradient guides refinement and provides a signal for stopping. To train EBRs effectively, we further introduce exploration-based training to discover additional modes and to better shape the energy landscape over a broader set of states. Across four algorithmic reasoning benchmarks, EBRs match or outperform prior recurrent reasoning models, achieving 99.9% accuracy on Sudoku-Extreme while using 345× less inference compute than the next-best model and maintaining comparable training compute. On all multi-solution benchmarks, EBRs achieve state-of-the-art coverage of distinct valid solutions while maintaining near-100% single-sample accuracy. These results suggest that energy-based models are a promising approach for strong reasoning performance.

Why Energy?

Existing looped reasoners update their latent state without any explicit assessment of it. An EBR learns an energy function that guides refinement, stopping and selection.

Energy, gradient norm and residual during descent, solved vs wrong

Correct solutions settle at low-energy stationary states. Energy, gradient norm and step size during descent on Sudoku-Extreme. Correct trajectories settle at low energy with small gradients, while wrong ones keep moving.

Test-Time Scaling

Running more descent steps refines each answer further, and running more trajectories gives the energy more candidates to choose the lowest-energy one from.

Sudoku-Extreme accuracy vs iterations with one sample (left), and vs number of samples (right)

EBRs reach EqR's best accuracy with 7× fewer iterations (left) and 43× fewer samples (right).

GRAM and TRM from GRAM's paper; EqR from its released checkpoint. Mean ± SD over three seeds.

Adaptive Stopping

Each trajectory stops once its energy gradient falls below a threshold, so easy puzzles finish in a few steps and hard ones keep going.

Sudoku-Extreme exact accuracy vs inference FLOPs per puzzle
Method# ParamsSudoku-Extreme ↑Maze-Hard ↑
Single sample (%)
TRM5–7M87.485.3
EqR5–7M93.0–
FPRM7M94.287.0
GRAM10M86.6–
Looped flows5–7M97.986.7
EBROurs1.8M99.487.4
Multiple samples (%), number of samples N in parentheses
PTRM5–7M98.8(100)86.7(100)
EqR5–7M99.8(128)–
GRAM10M96.1(20)–
Looped flows5–7M99.3(5)86.9(5)
EBROurs1.8M99.9(5)88.3(5)

EBRs give the best single-sample accuracy among reasoning models. Within only 5 passes, they reach 99.9% and 88.3%, ahead of methods that draw up to 128. Stopping each trajectory once its energy gradient falls below a threshold keeps 99.9% accuracy with 10.9× fewer FLOPs than a fixed-depth EBR, allowing us to surpass EqR with 345× fewer FLOPs.

Multiple Solutions

EBR acts like stochastic gradient descent and can capture many different solutions. Coverage captures the share of each input's distinct valid solutions that 20 samples find.

MethodN-QueensGraph Coloring
8×810×108-vertex10-vertex
Single samplePass@1 (%) ↑Conflicts ↓
TRM66.817.5109.3170.7
GRAM99.589.78.331.3
EBROurs10099.500
20 samplesCov@20 (%) ↑Cov@20 (%) ↑
TRM22.62.022.36.8
GRAM89.757.585.851.3
EBROurs98.192.088.654.5

Pass@1 (%) and Conflicts (edges summed over the test graphs) use one sample. Cov@20 is the share of distinct valid solutions found by 20 samples (%). Baselines from GRAM's paper.

Graph Coloring, 10 vertices

BibTeX

@misc{chen2026energybasedreasoners,
  title         = {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},
  year          = {2026},
  eprint        = {2611.08977},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2611.08977}
}
联系我们 contact @ memedata.com