Saved in:
| Main Authors: | , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2507.06607 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866918170909999104 |
|---|---|
| author | Ren, Liliang Chen, Congcong Xu, Haoran Kim, Young Jin Atkinson, Adam Zhan, Zheng Sun, Jiankai Peng, Baolin Liu, Liyuan Wang, Shuohang Cheng, Hao Gao, Jianfeng Chen, Weizhu Shen, Yelong |
| author_facet | Ren, Liliang Chen, Congcong Xu, Haoran Kim, Young Jin Atkinson, Adam Zhan, Zheng Sun, Jiankai Peng, Baolin Liu, Liyuan Wang, Shuohang Cheng, Hao Gao, Jianfeng Chen, Weizhu Shen, Yelong |
| contents | Recent advances in language modeling have demonstrated the effectiveness of State Space Models (SSMs) for efficient sequence modeling. While hybrid architectures such as Samba and the decoder-decoder architecture, YOCO, have shown promising performance gains over Transformers, prior works have not investigated the efficiency potential of representation sharing between SSM layers. In this paper, we introduce the Gated Memory Unit (GMU), a simple yet effective mechanism for efficient memory sharing across layers. We apply it to create SambaY, a decoder-hybrid-decoder architecture that incorporates GMUs in the cross-decoder to share memory readout states from a Samba-based self-decoder. SambaY significantly enhances decoding efficiency, preserves linear pre-filling time complexity, and boosts long-context performance, all while eliminating the need for explicit positional encoding. Through extensive scaling experiments, we demonstrate that our model exhibits a significantly lower irreducible loss compared to a strong YOCO baseline, indicating superior performance scalability under large-scale compute regimes. Our largest model enhanced with Differential Attention, Phi4-mini-Flash-Reasoning, achieves significantly better performance than Phi4-mini-Reasoning on reasoning tasks such as Math500, AIME24/25, and GPQA Diamond without any reinforcement learning, while delivering up to 10x higher decoding throughput on 2K-length prompts with 32K generation length under the vLLM inference framework. We release our training codebase on open-source data at https://github.com/microsoft/ArchScale. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2507_06607 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Decoder-Hybrid-Decoder Architecture for Efficient Reasoning with Long Generation Ren, Liliang Chen, Congcong Xu, Haoran Kim, Young Jin Atkinson, Adam Zhan, Zheng Sun, Jiankai Peng, Baolin Liu, Liyuan Wang, Shuohang Cheng, Hao Gao, Jianfeng Chen, Weizhu Shen, Yelong Computation and Language Machine Learning Recent advances in language modeling have demonstrated the effectiveness of State Space Models (SSMs) for efficient sequence modeling. While hybrid architectures such as Samba and the decoder-decoder architecture, YOCO, have shown promising performance gains over Transformers, prior works have not investigated the efficiency potential of representation sharing between SSM layers. In this paper, we introduce the Gated Memory Unit (GMU), a simple yet effective mechanism for efficient memory sharing across layers. We apply it to create SambaY, a decoder-hybrid-decoder architecture that incorporates GMUs in the cross-decoder to share memory readout states from a Samba-based self-decoder. SambaY significantly enhances decoding efficiency, preserves linear pre-filling time complexity, and boosts long-context performance, all while eliminating the need for explicit positional encoding. Through extensive scaling experiments, we demonstrate that our model exhibits a significantly lower irreducible loss compared to a strong YOCO baseline, indicating superior performance scalability under large-scale compute regimes. Our largest model enhanced with Differential Attention, Phi4-mini-Flash-Reasoning, achieves significantly better performance than Phi4-mini-Reasoning on reasoning tasks such as Math500, AIME24/25, and GPQA Diamond without any reinforcement learning, while delivering up to 10x higher decoding throughput on 2K-length prompts with 32K generation length under the vLLM inference framework. We release our training codebase on open-source data at https://github.com/microsoft/ArchScale. |
| title | Decoder-Hybrid-Decoder Architecture for Efficient Reasoning with Long Generation |
| topic | Computation and Language Machine Learning |
| url | https://arxiv.org/abs/2507.06607 |