1 paper · 1 filter
Lukas Fesser, Mozes Jacobs, Thomas Fel +2
When attention concentrates on a single token, a sink, what is the model actually computing? Attention sinks are ubiquitous in softmax transformers, yet this shared visual signatur…