The Necessary Sink

Attention sinks are a well-documented phenomenon in transformers: certain tokens — typically the first token in a sequence — receive disproportionately high attention scores across many layers, despite having no special semantic role. The token carries no more information than its neighbors. It just absorbs attention. The standard interpretation is that sinks are a pathological artifact of softmax normalization — the model needs to put its attention somewhere when nothing is relevant, and the first token becomes the default dump.

arXiv:2603.11030 (2026) proves that attention sinks are not pathological. They are necessary. For trigger-conditional tasks — where the model must produce different outputs depending on whether a specific trigger appeared earlier in the sequence — softmax transformers provably cannot compute the correct answer without concentrating excess attention on at least one token that serves no other role.

The mechanism is structural. Softmax attention normalizes: every position’s attention weights must sum to one. When the model needs to compute a trigger-conditional output, it must distinguish “trigger present” from “trigger absent” at every subsequent position. But the information about trigger presence is a single bit, and distributing a single bit across the full attention distribution of every layer is inefficient. The model needs a place to store the bit passively — a token that absorbs attention when the trigger is absent and releases it when the trigger is present. That token is the sink.

The sink functions as a conditional register. When the trigger fires, the model routes attention away from the sink toward the trigger token, changing the downstream computation. When the trigger doesn’t fire, attention stays on the sink, producing the default behavior. The excess attention is not wasted capacity. It is the encoding of the null condition — the explicit representation of “nothing happened.” The model can’t represent “nothing happened” by attending to nothing, because softmax requires positive weights everywhere. It represents “nothing happened” by attending to a designated parking location.

The general principle: when a normalization constraint forces all resources to be allocated (as softmax forces all attention to be distributed), the system must allocate resources even when there is nothing to attend to. The allocation to “nothing” becomes a signal — it encodes the absence of the thing the system is looking for. The artifact is the mechanism. What looked like wasted capacity was a necessary encoding of the null state, forced into existence by the same normalization that makes attention work in the first place.