attention sink

When you inspect where a transformer's attention actually goes, a surprising amount of it piles onto the very first tokens of the sequence, often a near-meaningless beginning-of-text token. These are attention sinks: positions that soak up attention weight not because they are informative but because softmax forces the attention distribution to sum to one and the model needs somewhere to dump the mass when there is no strong match.

The mechanism follows from softmax being a normalized exponential: every head must distribute a full unit of attention even when no key is relevant, so it learns to park the excess on a few fixed early positions whose keys it has calibrated as a default. The practical consequence appears in streaming generation: if you naively evict old tokens with a rolling window to bound the KV cache, dropping those first few sink tokens destroys the calibrated normalization and quality collapses. StreamingLLM showed that simply keeping a handful of initial sink tokens cached alongside the recent window restores stable generation over millions of tokens.

Attention sinks explain why sliding-window caches must retain the first tokens, and they connect to why some architectures now add a dedicated learnable sink token so the model has an explicit no-op target rather than commandeering a content token. They are also a lens on quantization and interpretability: sink positions carry outsized activation magnitudes, which produces the massive-activation outliers that make low-bit quantization hard and that show up repeatedly when probing what individual heads actually do.