[论文] Classical Energies Learn to Reason 1,000x Faster than Looped Models
研究领域: ML 作者: Benjamin Hoover, Hugh Blayney, James C. R. Whittington 发布时间: 2025-11-13 arXiv: 2511.10432
论文概要
研究领域: ML 作者: Benjamin Hoover, Hugh Blayney, James C. R. Whittington 发布时间: 2025-11-13 arXiv: 2511.10432
中文摘要
循环模型已成为推理的突出范式,但可能不稳定、训练慢、且为简单谜题消耗大量参数与 FLOP。对能量函数下降则带来内置提前停止、随问题难度自适应的计算量、无需对求解器微分的快速训练,以及学习谜题约束的可解释权重。我们提出 CRAM——80 年代 Hopfield & Tank 能量的学习版本——在单 GPU 上训练 42ms 和 132ms 后分别解出 99.996% 的数独-极限和 100% 的迷宫-唯一测试题,比循环模型快至少 1,000 倍,在匹配精度下推理 FLOP 至少减少 282 倍。本工作表明:能量函数是从数据训练推理的极有前景的方向;同时警示不要用轻易可解的基准来测试循环推理技术的可扩展性。
原文摘要
Looped models have emerged as a prominent paradigm for reasoning, but they can be unstable, slow to train, and use many parameters and FLOPs. Descent on an energy offers built-in early stopping, compute that adapts to problem difficulty, fast training without differentiating through a solver, and interpretable weights. We demonstrate with CRAM, a learned version of Hopfield & Tank's energy from the 80s, which solves 99.996% of Sudoku-Extreme and 100% of Maze-Unique test puzzles after 42 ms and 132 ms of training on one GPU — at least 1,000x faster than looped models — using at least 282x fewer inference FLOPs at matched accuracy. Our work suggests energies are an extremely promising direction for training reasoning from data, and cautions against using trivially solvable benchmarks to test...
*自动采集于 2026-11-15*
#论文 #arXiv #ML #小凯