High Context window Question 154 of 223

What are attention sinks and why do they appear in streaming long context?

GenAI / LLM · Speak this in 60–90 seconds · Faridabad & Delhi NCR

PICTURE THIS: A SENTENCE BECOMES TOKENS

The model does not read letters like humans. It reads these pieces, then predicts the next one.

Simple meaning

Models often park large attention mass on the first few tokens.

WHY — Tokens instead of words?

LLMs use tokens (not full words) because it helps them:

Context window questions separate

people who only read docs from people who shipped.

Keep it short, concrete,

and tied to GenAI / LLM work.

Stable token IDs

Each piece maps to a number the network can learn.

Fits the model

Fixed pieces are what transformers expect as input.

STEPS — What happens with tokens?

Before the model can read a sentence, it goes through these steps:

1
Models often park large
attention mass on the first few tokens.
2
Streaming or sliding-window schemes
keep those sink tokens plus a recent window so quality does not fall off a cliff.
3
It is an inference
hack around the fact that softmax attention wants somewhere to put residual mass.
4
Context mix
Attention looks at nearby tokens together.
5
Next token
The model scores what should come next.
6
Decode
IDs turn back into readable text.

EXAMPLE — See it in action

Let's see how a real sentence is tokenized (tokens may vary by model):

Input text

“Streaming or sliding-window schemes keep those sink tokens plus a recent window ”

Tokenized output

Streamingorslidingwindowschemeskeep

Token IDs (example)

2987408337471632900

Note: Actual tokens and IDs depend on the tokenizer (e.g., GPT, Llama, etc.).

Key takeaway

Models often park large attention mass on the first few tokens. Streaming or sliding-window schemes keep those sink tokens plus a recent window so quality does not fall off a cliff.