Teortaxes proposes affinity-based routing for weights
Reply suggests letting weight blocks learn mutual preferences to boost specialization and expressivity.
Teortaxes replied to eliebakouch suggesting a routing algorithm that lets weight blocks learn mutual affinities and vote preferences across layers. The proposal aims for greater expressivity and specialization. Other visible replies reference depth-wise batching from papers on recursive models, recurrent depth computation, and past rumors around looped transformers in models like Mythos. One user noted trainability issues with high forward pass counts.
Combined views
60.1K
15 posts, first seen 21h ago