Decoder-Hybrid-Decoder Architecture for Efficient Reasoning with Long Generation

TL;DR

This paper introduces SambaY, a hybrid decoder architecture with shared memory mechanisms that significantly improves long-context reasoning efficiency and scalability in language models, outperforming previous models in speed and accuracy.

Contribution

The paper proposes the Gated Memory Unit (GMU) and SambaY architecture, enabling efficient memory sharing and improved long-context reasoning without positional encoding.

Findings

SambaY achieves up to 10x higher decoding throughput on long prompts.

Model exhibits lower irreducible loss and better scalability than YOCO baseline.

Significant performance improvements on reasoning benchmarks like Math500 and AIME24/25.

Abstract

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.