LLM Attention Visualization

Transformer models can pull from any earlier token while generating, and this page makes that selection visible. A model of about 600 million parameters writes a reply in your browser through Transformers.js, and tapping or hovering on any output token fades the earlier text according to how much it contributed.

The number behind each fade is a heavy simplification: attention weights scaled by the value vector's magnitude, combined across all heads and summed over every layer, then normalized so the strongest token is fully opaque. Patterns still show up, such as addresses and dates copied verbatim in the sample office-move summary lighting up their source lines.