Attention sink
Transformer models dump a disproportionate share of their attention on the very first tokens of the sequence, regardless of what those tokens mean.
The phenomenon was described in 2023 and has a simple explanation: softmax forces the attention weights to sum to one, so even when no position is relevant the mass has to go somewhere. During training the model learns to dump the excess on the opening tokens, which every sequence has and which are therefore always available. The consequence is practical and surprising: evict those first tokens from the KV cache to save memory — which is exactly what a sliding window does — and the model's output collapses into nonsense, even though the tokens in question were usually just a newline or a system prompt. Keeping a handful of them fixes it completely.