Hybrid sequence model paper argues access structure can outweigh scale

A new arXiv paper argues that capability in hybrid sequence models can come from access structure, not only from model scale. The authors propose lower bounds and pre-registered tests to examine the relationship.

The question matters because AI progress is often discussed in terms of parameter count and compute. If architecture and access patterns can change capability in meaningful ways, evaluation needs to look beyond size alone.

The paper adds to research on how model design choices shape what systems can learn, remember, and generalize.