What changes a token state inside one transformer block?

Read context, transform locally, keep what helps

A model passes every token state through a stack of transformer blocks. Each block has two complementary jobs. Attention lets a position gather relevant information from other permitted positions. The MLP then transforms the information available at each position independently. Residual connections add each result to a shared working state instead of replacing it, so useful information can persist while new evidence accumulates. Normalization keeps repeated updates manageable. One block usually makes only a partial refinement; depth lets many refinements compose. Think of the stack as a team repeatedly annotating the same working document, not an assembly line where each layer owns one tidy linguistic rule.

Attention communicates between positions; the MLP transforms each position.