Because Large Language Models use subword tokenization (like BPE), a word like indubitably might be split into in, du, bit, ably. If you try to visualize the raw attention matrix, you get a massive Token × Token grid that looks like a noisy, unreadable mess to human eyes.

To build a human-readable Word-Level Attention Map, we have to mathematically aggregate those token weights.

If Word A is made of 3 tokens, and Word B is made of 2 tokens, how much does Word A attend to Word B? We calculate the “attention from Word A to Word B” by taking the mathematical average of the attention weights going from all 3 sub-tokens of A to all 2 sub-tokens of B.

It’s a mathematical compromise, but it collapses the massive token grid into a clean Word × Word 2D matrix that we can actually interpret.