Attention Sink

Creator

Creator

Seonglae Cho

Created

Created

2023 Oct 14 12:48

Editor

Editor

Seonglae Cho

Edited

Edited

2024 Nov 27 15:4

Refs

Refs

Continual Learning

Sparse Attention

Vertical strong line to
System Prompt or
BOS Token (>50%) in
Attention Score

LLMs aggregate information through Pyramidal Information Funneling where attention is scattering widely in lower layers and ultimately focusing on critical tokens.

SINK TOKEN

Even though keeping the first token might seem semantically meaningless, it has significance. This is because due to the characteristics of the Attention Mechanism, the first token is used as an anchor for calculating the Attention Score through positional embedding. Therefore, even if it's semantically meaningless, the model structurally requires it.

Attention Sinks Notion

REGISTERS Tokens

Attention Canceling

Attention sink

https://arxiv.org/pdf/2402.09221

Efficient Streaming Language Models with Attention Sinks

Deploying Large Language Models (LLMs) in streaming applications such as multi-round dialogue, where long interactions are expected, is urgently needed but poses two major challenges. Firstly,...

Efficient Streaming Language Models with Attention Sinks

https://arxiv.org/abs/2309.17453

Efficient Streaming Language Models with Attention Sinks

🕳️ Attention Sinks in LLMs for endless fluency

A Blog post by Tom Aarsen on Hugging Face

🕳️ Attention Sinks in LLMs for endless fluency

https://huggingface.co/blog/tomaarsen/attention-sinks

🕳️ Attention Sinks in LLMs for endless fluency

Null attention

https://arxiv.org/pdf/1906.04284

Massive Activation

Special token, delimiter, conjunction, preposition, first token, number token, weak semantics

https://arxiv.org/pdf/2402.17762

When it emerges

https://arxiv.org/pdf/2410.10781

Recommendations

//////////