arXiv:2611.23508v1 Announce Type: new Abstract: Large vision-language models (LVLMs) and language models (LLMs) often exhibit attention sinks, whereby particular positions, such as the beginning-of-sequence (BOS) token, receive a disproportionate share of attention. Under standard causal masking, however, the BOS query attends only to itself, so the representation that later tokens repeatedly acce
Straipsnis generuojamas iš faktų karkaso pasirinktu stiliumi — be išgalvojimų.
Paskelbta: 2026-11-20 05:00