titleDepthGraft: Structural Regularization through Hierarchical Cross-Layer KV Reconstruction
authorsXiaohan Qin, Xiangdong Zhang, Yu Wang, Huaijin Wu, Zhuo Xia, Yebin Yang, Junchi Yan
date2026-11-02
venuearXiv (preprint, not peer reviewed)
idarXiv:2611.03142
affiliationdots studio
upvotes15 (Hugging Face Daily Papers, 2026-11-13)
linksabs pdf hf code
cite
relatedDoSA: Dots Sparse Attention
xHC: Expanded Hyper-Connections
Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information
On the Design of Hyper-Connections: Simplification, Stability, and Scalability
VibeSearchBench: Benchmarking Long-horizon Proactive Search in the Wild
abstractScaling 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 KV representation geometry. We provide a systematic analysis of this mechanism, showing that proper sharing topologies sharpen key retrieval directions while enriching the value content subspace, whereas existing adjacent or global-anchor designs do not fully exploit this effect. Motivated by this, we propose DepthGraft, a hierarchical KV-sharing architecture that reconstructs deep-layer KV from stride-aligned shallow-middle source pairs. DepthGraft distributes non-local gradient flow across source layers, while preserving prefill early completion and reducing KV cache by roughly one third. Experiments on dense and MoE models from 0.5B to 16B parameters show consistent downstream improvements under both GQA and MLA. Notably, on the 16B MoE model, DepthGraft achieves a 7.0% absolute improvement on MMLU-Pro, along with gains of 6.6% on CMMLU and 6.4% on C3. These results validate the potential of DepthGraft as a scalable design for large-scale pretraining. The project page is available at https://github.com/jianke0604/DepthGraft.

bibtex

@misc{qin2026depthgraft,
  title={DepthGraft: Structural Regularization through Hierarchical Cross-Layer KV Reconstruction},
  author={Xiaohan Qin and Xiangdong Zhang and Yu Wang and Huaijin Wu and Zhuo Xia and Yebin Yang and Junchi Yan},
  year={2026},
  eprint={2611.03142},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2611.03142}
}