Uloženo v:
Podrobná bibliografie
Hlavní autor: Lirov, Kharki
Médium: Recurso digital
Jazyk:angličtina
Vydáno: Zenodo 2026
Témata:
On-line přístup:https://doi.org/10.5281/zenodo.19840136
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Obsah:
  • <p>We introduce Motif-Upcycling, a structure-preserving framework for adapting pretrained Transformer models. The key idea is that common feed-forward modules, including SwiGLU FFNs, can be exactly factorized along their intermediate channel axis into motif-aligned components. With neutral routing, the factorized module computes the same function as the original pretrained block at initialization. We further introduce Scale-Aware Residual Control (SARC), an identity-preserving control motif that modulates the magnitude of trainable residual interventions relative to the residual stream. We also propose Emergence as Coupled Budget Thresholds (ECBT), a conditional model showing that apparent capability cliffs can arise from multiplicatively coupled motif effectiveness curves under uniform budget allocation.</p>