arXiv cs.CL· Xiaohan Qin, Xiangdong Zhang, Yu Wang, Huaijin Wu, Zhuo Xia, Yebin Yang, Junchi Yan·· 11 小时前AI 评分34
DepthGraft：通过分层跨层 KV 重建实现结构正则化
DepthGraft: Structural Regularization through Hierarchical Cross-Layer KV Reconstruction
AI 导读
DepthGraft 是一种分层 KV 共享架构，从步幅对齐的浅层-中层源对重建深层 KV，在保持 prefill 提前完成的同时将 KV cache 减少约三分之一。在 0.5B 至 16B 参数的稠密与 MoE 模型上，GQA 和 MLA 下均取得一致的下游提升；16B MoE 模型上 MMLU-Pro 绝对提升 7.0%，CMMLU 提升 6.6%，C3 提升 6.4%。
正文 · 原文
Abstract: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 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 this https URL.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2611.03142 [cs.CL] |
| (or arXiv:2611.03142v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2611.03142 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Xiaohan Qin [view email]
[v1]
Mon, 2 Nov 2026 13:52:03 UTC (9,235 KB)
来源：arXiv cs.CL · arxiv.org