2026-11-11 · America/Los_Angeles · 论文 · #7

DepthGraft: Structural Regularization through Hierarchical Cross-Layer KV Reconstruction

Scaling depth is a key route to stronger large language models, but Pre-LN architectures often suffer from depth-wise information dilution, limiting the effective use of deep layers. While cross-layer connectivity has proven effective at mitigating this problem, existing designs usually introduce additional computation or memory access. We revisit cross-layer KV sharing as an efficient form of such connectivity. Beyond cache compression, each KV reuse edge creates a direct gradient path from deep layers to source KV projections, turning KV sharing into an implicit structural regularizer for…

热度 43.4 / 100；排名与评分保留该期记录。