From our lab.
MANUSCRIPTS / NOTES / PROPOSALSHidden State Emulation
A careful investigation of sparse Green operators, local neural approximations, and the gap between a good hidden-state fit and a safe replacement.
Manuscript · exploratory resultsContinued-Fraction Gated Networks
An architecture note drawn from our uploaded character-model implementation: gated rational updates, multiscale carry, and latent attention.
Implementation note · historical metricGrowing Intelligence, Bounded Compute
Can a graph expand its stored capabilities while keeping each training step compact and regular? A testable direction, with the open questions left visible.
Proposal · not benchmarkedTHE READING DESK
Ideas worth
thinking with.
Selected external papers informing our questions. These works belong to their authors.
On Efficient Scaling of GNNs via IO-Aware Layers Implementations
A useful reminder that sparse graph arithmetic is only part of the cost. Memory movement and edge intermediates also determine training speed.
Read the original paperRevisiting Pre-Propagation GNNs
Explores diffusion operators and hidden-state re-propagation: relevant to our question of separating expensive graph work from routine training.
Read the original paperExpander Graph Propagation
A starting point for thinking about sparse communication paths that still connect distant parts of a system.
Read the original paper