论文卡-Attention Is All You NeedPaper Card - Attention Is All You Need
论文卡 · 约 7 分钟Paper card · about 7 min · 引用核验 2026-11-03（仅引用链接与日期） · refs checked 2026-11-03 (reference links and dates only)
Transformer 的源头论文。这张卡列清它要解决什么、证据在哪一节、以及它没有主张什么。The source paper of the Transformer. This card lays out the question it set out to solve, where the evidence lives, and what the paper does not claim.
研究问题：只用注意力、不用循环和卷积，能不能做序列转换——摘要原话 "dispensing with recurrence and convolutions entirely"（完全弃用循环与卷积）。
核心证据：第 4 节与 Table 1 对比三种架构量——最大路径长度、每层计算量、最少顺序操作数；实验是两个机器翻译任务，WMT 2014 英德翻译 28.4 BLEU（摘要原话）。
作者自陈局限：这篇论文没有独立的 Limitations 节（当日核对 arXiv HTML 版）；最接近的是第 7 节注意力可视化——那是机制解读，不是局限声明。适用范围要读者自己补判断。
它没有主张的：论文没有说这套架构能训出通用语言模型。「Transformer 是大模型的地基」是本站回顾性说法，不是论文结论。
读法提示：Table 1 的三量对比才是论证核心；只看架构图会错过论文真正的主张。
Research question: can sequence transduction work with attention alone — the abstract's "dispensing with recurrence and convolutions entirely".
Core evidence: Section 4 and Table 1 compare three architectures on path length, per-layer compute, and sequential operations; experiments are two machine-translation tasks, 28.4 BLEU on WMT 2014 EN-DE (abstract).
Author-stated limitations: the paper has no standalone Limitations section (verified against the arXiv HTML version on 2026-10-29); Section 7's attention visualizations explain the mechanism, not the limits — applicability is the reader's job.
What it does not claim: the paper never says this architecture will train general-purpose language models. "Transformer is the foundation of LLMs" is our retrospective framing, not the paper's conclusion.
Reading tip: the three-way complexity comparison in Table 1 is the actual argument; looking only at the architecture diagram misses the claim.
例子与判断Example and interpretation
以下是教学情境，不是本站运行或实测记录。A teaching scenario, not a measured or executed site test.
带着“没有循环结构，序列中的位置和依赖如何表示”去读方法节。将位置编码、注意力和训练设置分别记下，再看翻译任务的实验；不要从单一任务结果推断所有后续模型的表现。Read the method asking how positions and dependencies are represented without recurrence. Separate positional encoding, attention and training setup, then inspect translation experiments. Do not generalize those results to every later model.