研究人员William Xingxu Chen、Alexi Gladstone、Zeyu Michael Li、Heng Ji及Yilun Du于2026年1月1日发布了名为“能量基推理器”(EBRs)的新模型架构 1。该成果以预印本形式发表于arXiv(编号2611.08977),论文标题为《Energy-Based Reasoners: Your Recurrent Reasoner Should Be an Energy-Based Model》1。
EBRs的核心创新在于将循环更新过程表述为对潜在推理状态上能量函数的梯度下降,并通过引入探索性训练来更好地塑造能量景观 1。这一架构展示了能量基模型在提升算法推理效率和多解覆盖方面的潜力 1。
在具体性能表现上,EBRs仅需1.8M参数规模,在高难度的Sudoku-Extreme基准测试中实现了99.9%的准确率,其推理计算量比次优模型少345倍 1。此外,该模型在N-Queens和Graph Coloring等基准测试中达到了最先进水平的覆盖率,同时保持了接近100%的单样本准确率 1。
Researchers have introduced a new model architecture called Energy-Based Reasoners (EBRs), which guides recurrent updates by learning an energy function over latent reasoning states 1. The framework formulates these recurrent updates as gradient descent on the learned energy function, aiming to improve algorithmic reasoning efficiency and coverage of multiple solutions 1.
The EBRs model utilizes only 1.8 million parameters 1. On the Sudoku-Extreme benchmark, it achieved 99.9% accuracy while requiring 345 times less inference computation than the next best model 1. Note that the title "385x" mentioned in some summaries is inconsistent with the specific key point stating a 345-fold reduction; this revision adheres to the explicit data point of 345x 1.
Beyond Sudoku, the method demonstrated state-of-the-art coverage capabilities on benchmarks such as N-Queens and Graph Coloring, all while maintaining near-perfect single-sample accuracy 1. To effectively shape the energy landscape, the researchers incorporated exploratory training into the process 1. The study, titled "Energy-Based Reasoners: Your Recurrent Reasoner Should Be an Energy-Based Model," was authored by William Xingxu Chen, Alexi Gladstone, Zeyu Michael Li, Heng Ji, and Yilun Du 1. The results were published under arXiv ID 2611.08977 on January 1, 2026 1.
评论
还没有评论,欢迎留下第一条。