讨论串评论krackers · 2025-05
“Turns out that https://www.fortressofdoors.com/four-magic-words/ was right and all you need to do in training is have the LLM meditate on a single example.”
用单个样本做 RLVR,就能大幅提升模型的数学推理表现
论文表明,仅用一个训练样本做带可验证奖励的强化学习(1-shot RLVR),就能有效激发 LLM 的数学推理能力。在 Qwen2.5-Math-1.5B 上,一个样本把 MATH500 从 36.0% 提到 73.6%,六个数学基准的平均值从 17.6% 提到 35.7%,与使用 1.2k 条 DeepScaleR 子集的结果相当;两个样本的结果还略高。作者还观察到跨类别泛化、训练准确率饱和后测试性能仍持续提升(post-saturation generalization),并验证其主要来自策略梯度损失、与 grokking 不同。
mentions 中没有可引用的评论原话,只能确定它在不同月份被多人提到,原因无法据此判断。
“Turns out that https://www.fortressofdoors.com/four-magic-words/ was right and all you need to do in training is have the LLM meditate on a single example.”