ISpring 2026
Knox theorem
Given a task that decomposes into independent sub-tasks, a system of small models specialized to each sub-task returns more correct answers than a single general model of equal aggregate parameter count. The theoretical foundation behind YG3 platform orchestration.
Read the paperII2026
Multiplicative Gate Adapters for Looped Transformers
Adapting looped transformer architectures with multiplicative gate adapters. Full paper archived on Zenodo.
Read the paper