PART 01
独立投影Separate projections
每个 head 有自己的一套 Q、K、V 投影。Each head owns a separate Q, K, and V projection.
01 · 同时看三种线索01 · Track three clues at once
读“女孩把书给男孩，因为他需要它”时，你会只盯着一种关系吗？In “The girl gave the boy a book because he needed it,” would one relationship be enough?
02 切换三个 head，再拖动合并比例Switch Heads, Then Adjust Their Merge
3 路并行3 parallel views
当前视图Inspected view
HEAD 1
指代关系reference relation
最粗来源Thickest source
书
48%
合并主导Merge leader
H1
45%
S3 线宽语法不变：lineWidth = 1.5 + weight × 10。切换 head 改变“哪套权重”，拖合并比例改变最终混合。S3 line grammar stays fixed: lineWidth = 1.5 + weight × 10. Switching heads changes the weight set; merge sliders change the final mix.
03 多头注意力的 4 个零件Four Parts of Multi-Head Attention
split · attend · joinPART 02
并行权重Parallel weights
各 head 对同一序列独立分配注意力。Heads assign attention over the same sequence independently.
PART 03
各自结果Per-head results
每一路先形成自己的加权汇总。Each path first produces its own weighted result.
PART 04
合并投影Joined projection
结果拼接后再投影回统一表示。Results concatenate, then project back into one representation.
04 最小机制：先各看各的，再装回一个向量Minimal Mechanism: Separate Views, Then One Vector Again
Concat(head1…headh)WO
headi = Attention(QWiQ, KWiK, VWiV)
三个滑杆是教学用合并比例。真实 Transformer
通常用学习到的输出投影完成合并，不是人工命名每个 head。The sliders are a teaching proxy for combination. A real Transformer usually learns an output projection;
people do not manually name every head.
复制成多路Fan into heads同一输入，不同投影Same input, different projections
并行注意Attend in parallel每路得到一套线宽Each path gets its own widths
合并输出Merge outputs多种关系进入同一表示Views return to one representation
05 边界实验：让一路吞掉全部合并权重Boundary Test: Let One Head Swallow the Entire Merge
主动制造失败Create a failure有多个 head，不等于真的使用了多种关系Having many heads does not mean many views survive
把 Head 1 拉到 100%，其余接近 0。并行结构仍在，但合并结果几乎只剩一路。Push Head 1 to 100% and the others near zero. Parallel paths still exist, but the merged result keeps almost one view.
尚未执行。三路都会进入合并结果。Not run yet. All three views enter the merged result.
✗ 一个 head 就是一个 Agent。One head is one Agent.
→ ✓ head 是一次并行注意力计算路径。A head is one parallel attention path.
✗ 每个 head 必然有清楚语义。Every head must have a clear meaning.
→ ✓ 角色标签只是教学视图，不是保证。Role labels are teaching views, not guarantees.
✗ head 越多一定越聪明。More heads always mean more intelligence.
→ ✓ 数量、维度与任务效果需要一起验证。Count, dimension, and task results must be evaluated together.
06 · 一句话带走06 · One line to keep