以下为上游提供的源文稿,可能存在识别错误。
乔木转写是基于源文稿的中文提炼,不是逐字翻译。请核对原文中的人名、数字与观点。
选一篇,慢慢读
从左侧挑一篇,正文会在这里排成书页。
When high-quality data runs out, a common practice is to train on it for several epochs, but each repeat is worth less than the last. Latent feedback, which feeds each token's top-layer state into the next token's input, is another way to spend that compute, yet it has not been tested against a vanilla Transformer that
阅读原文 · Benhao Huang, Chaitanya Dwivedi, Xuezhe Ma, Alexander Moreno, Hector Liu
向阳乔木 · opencode-go/deepseek-v4.1-flash
当高质量数据快用完时,大家通常会拿同一批数据多训几轮。问题是每多训一轮,收益就衰减一次。这篇论文问了一个很实际的问题:同样一笔算力,是拿去多训几轮,还是换一种架构设计? Latent feedback 是这里的主角。做法是把每个 token 顶层输出的状态,喂回下一个 token 的输入。听起来像个小改动,但作者把它放到了和“多训几轮”正面对比的位置上,而且...