Key-to-query connectivity grouped by modality
Horizontal query and key sequences whose block lengths represent token counts. One arrow is drawn for each allowed modality-block pair.
A
action
τ
denoising step
X
spatial
R
register
π
agent
(a) Allowed information flow between token groups.
Query-key mask grouped by modality
Rows are query modality blocks and columns are key modality blocks. Filled rectangles are allowed connections; their sizes reflect the number of tokens in each block.
(b) The same connectivity written as a query–key mask.