Research

Paper

AI LLM March 16, 2026

Mixture-of-Depths Attention

Authors

Lianghui Zhu, Yuxin Fang, Bencheng Liao, Shijie Wang, Tianheng Cheng, Zilong Huang, Chen Chen, Lai Wei, Yutao Zeng, Ya Wang, Yi Lin, Yu Li, Xinggang Wang

Abstract

Scaling depth is a key driver for large language models (LLMs). Yet, as LLMs become deeper, they often suffer from signal degradation: informative features formed in shallow layers are gradually diluted by repeated residual updates, making them harder to recover in deeper layers. We introduce mixture-of-depths attention (MoDA), a mechanism that allows each attention head to attend to sequence KV pairs at the current layer and depth KV pairs from preceding layers. We further describe a hardware-efficient algorithm for MoDA that resolves non-contiguous memory-access patterns, achieving 97.3% of FlashAttention-2's efficiency at a sequence length of 64K. Experiments on 1.5B-parameter models demonstrate that MoDA consistently outperforms strong baselines. Notably, it improves average perplexity by 0.2 across 10 validation benchmarks and increases average performance by 2.11% on 10 downstream tasks, with a negligible 3.7% FLOPs computational overhead. We also find that combining MoDA with post-norm yields better performance than using it with pre-norm. These results suggest that MoDA is a promising primitive for depth scaling. Code is released at https://github.com/hustvl/MoDA .

Metadata

arXiv ID: 2603.15619
Provider: ARXIV
Primary Category: cs.CL
Published: 2026-03-16
Fetched: 2026-03-17 06:02

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.15619v1</id>\n    <title>Mixture-of-Depths Attention</title>\n    <updated>2026-03-16T17:59:55Z</updated>\n    <link href='https://arxiv.org/abs/2603.15619v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.15619v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Scaling depth is a key driver for large language models (LLMs). Yet, as LLMs become deeper, they often suffer from signal degradation: informative features formed in shallow layers are gradually diluted by repeated residual updates, making them harder to recover in deeper layers. We introduce mixture-of-depths attention (MoDA), a mechanism that allows each attention head to attend to sequence KV pairs at the current layer and depth KV pairs from preceding layers. We further describe a hardware-efficient algorithm for MoDA that resolves non-contiguous memory-access patterns, achieving 97.3% of FlashAttention-2's efficiency at a sequence length of 64K. Experiments on 1.5B-parameter models demonstrate that MoDA consistently outperforms strong baselines. Notably, it improves average perplexity by 0.2 across 10 validation benchmarks and increases average performance by 2.11% on 10 downstream tasks, with a negligible 3.7% FLOPs computational overhead. We also find that combining MoDA with post-norm yields better performance than using it with pre-norm. These results suggest that MoDA is a promising primitive for depth scaling. Code is released at https://github.com/hustvl/MoDA .</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CL'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AI'/>\n    <published>2026-03-16T17:59:55Z</published>\n    <arxiv:comment>Code is released at https://github.com/hustvl/MoDA</arxiv:comment>\n    <arxiv:primary_category term='cs.CL'/>\n    <author>\n      <name>Lianghui Zhu</name>\n    </author>\n    <author>\n      <name>Yuxin Fang</name>\n    </author>\n    <author>\n      <name>Bencheng Liao</name>\n    </author>\n    <author>\n      <name>Shijie Wang</name>\n    </author>\n    <author>\n      <name>Tianheng Cheng</name>\n    </author>\n    <author>\n      <name>Zilong Huang</name>\n    </author>\n    <author>\n      <name>Chen Chen</name>\n    </author>\n    <author>\n      <name>Lai Wei</name>\n    </author>\n    <author>\n      <name>Yutao Zeng</name>\n    </author>\n    <author>\n      <name>Ya Wang</name>\n    </author>\n    <author>\n      <name>Yi Lin</name>\n    </author>\n    <author>\n      <name>Yu Li</name>\n    </author>\n    <author>\n      <name>Xinggang Wang</name>\n    </author>\n  </entry>"
}