Post

What do LLMs see? I wrote a lil' tool that extracts the attention matrices out of open models and creates this typing visual, with each token's opacity changing according to its average attention score as the prompt progresses. Dimmer words are considered less important to the model.

Andreas Møller

@andreasmoller.dk

So they are narcissists.

LLMs commonly use the first token as an attention "sink", giving it a high attention score regardless of what token it is. arxiv.org/abs/2504.02732

arxiv.orgWhy do LLMs attend to the first token?Large Language Models (LLMs) tend to attend heavily to the first token in the sequence -- creating a so-called attention sink. Many works have studied this phenomenon in detail, proposing various ways...

Also highly recommended read in this context ->

lesswrong.comThe Curious Case of the bos_token — LessWrongEdit: Simple application of these ideas achieved the GPT2 Speed Running world record of 171 seconds on 8H100 GPUs. https://github.com/KellerJordan/mo…

19/2/26, 7:43 pm0 reposts1 like

No replies yet.