Concept
Block-causal
Block-causal describes an attention masking scheme that sits between fully bidirectional and strictly left-to-right (token-level causal) attention. Think of standard causal attention, except the unit of ordering is a block of positions rather than a single token: inside a…
The rest of “Block-causal” is a premium feature: every concept in the library gets a precise, practitioner-focused write-up like this one, cross-linked straight from the paper summaries that use it.
Log in to unlock→