Why a model floods its first word with attention it never reads. · Source

It floods its very first word with attention it never reads.

Watch where a model looks, and you'll catch it pouring a flood of attention onto its first word — a word that often carries almost nothing. Not a glitch. Like a lightning rod: it installs a harmless spike to take the strike, so the storm passes through without wrecking the rest. The model keeps a place to dump attention it doesn't need.

Softmax must hand out a whole unit — even to nothing.

\alpha_j=\dfrac{e^{s_j}}{\sum_{k=1}^{t}e^{s_k}},\qquad \sum_{j=1}^{t}\alpha_j=1,\quad \alpha_j>0

Attention runs through softmax, which turns raw scores into weights that are all positive and sum to exactly one. So every word must spend a full unit of attention on the words it can see — even when none of them matter. Like a watering can you must tip empty: every last drop has to land somewhere, even on soil that's already soaked.

So it builds a drain for the attention it can't use.

If a word must spend a full unit but has nothing worth looking at, where does the surplus go? The model's fix: pick one spot and make it a drain. Like the overflow in a basin: it quietly carries off whatever the bowl can't hold, so nothing floods. That dumping-ground is the attention sink — and it's exactly what gives the trick its name.

Which word becomes the drain? The very first one.

\mathcal{V}_t=\{1,2,\dots,t\},\qquad \bigcap_{t=1}^{n}\mathcal{V}_t=\{1\}

Why the first word? Because of how a model reads — each word may look only at the words before it. That makes the very first token the one key every later word can always see. Like a tree's single trunk: follow any branch back far enough and you reach the same root. The first word is the one place all of them can reach — the natural drain.

Dumping there is harmless — it gives almost nothing back.

o_t=\sum_{j=1}^{t}\alpha_{tj}\,v_j,\qquad \alpha_{t1}\ \text{large},\ \ v_1\ \text{carries little}\ \Rightarrow\ \text{near no-op}

Why doesn't all that wasted attention corrupt the answer? Because a word's output is a blend of the values it attends to — and the sink's value carries almost no signal. Pour attention in and barely anything comes out. Like a wall of studio foam: shout at it and no echo returns. The sink absorbs the surplus and changes nothing.

Stream a long chat, and you must never evict the sink.

\underbrace{\{1,\dots,k\}}_{\text{sinks}}\ \cup\ \underbrace{\{t-w+1,\dots,t\}}_{\text{recent window}}\qquad (k\approx 4)

To keep memory bounded, a model can slide a window and forget its oldest words. But forget the sink, and the output collapses into nonsense. The cure is simple: always keep the first few tokens, plus the recent window. Like the keystone of an arch: every other stone can be swapped, but pull that one and the whole span falls.

The cleaner fix: let attention add up to less than one.

\alpha_j=\dfrac{e^{s_j}}{1+\sum_{k}e^{s_k}},\qquad \sum_{j}\alpha_j=\dfrac{\sum_{k}e^{s_k}}{1+\sum_{k}e^{s_k}}<1

What if a word could simply hold attention back? Add a single +1 to softmax's denominator — an invisible slot scored zero — and the weights are free to sum to less than one. Now a word can attend to nothing without inventing a drain. Like a spillway beside a dam: a channel built on purpose to send the surplus harmlessly nowhere.

It made a place for what it didn't need. Was that a flaw?

We almost called the sink a bug — a word drowning in attention it never reads. But it was the model quietly building somewhere to put nothing, so the rest stayed clean. 🌱 How many flaws are really a mind making room — leaving an empty space that simply means none?