1 paper · 1 filter
Andrii Balashov, Olena Ponomarova
Conditional computation can decouple language model quality from per-token inference cost, yet leading techniques act on a single axis in isolation: Mixture-of-Experts (MoE) sparsi…