Paper · arXiv 2404.07904
HGRN2: Gated Linear RNNs with State Expansion
Zhen Qin, Songlin Yang, Weixuan Sun, Xuyang Shen, Dong Li, Weigao Sun, Yiran Zhong
The introduction of an outer-product-based state expansion mechanism in hierarchical gated linear RNN (HGRN) enhances its performance and expressiveness without additional parameters, outperforming architectures like Mamba and LLaMa in various tasks.
Abstract
Hierarchically gated linear RNN (HGRN,Qin et al. 2023) has demonstrated competitive training speed and performance in language modeling, while offering efficient inference. However, the recurrent state size of HGRN remains relatively small, which limits its expressiveness.To address this issue, inspired by linear attention, we introduce a simple outer-product-based state expansion mechanism so that the recurrent state size can be significantly enlarged without introducing any additional parameters. The linear attention form also allows for hardware-efficient training.Our extensive experiments verify the advantage of HGRN2 over HGRN1 in language modeling, image classification, and Long Range Arena.Our largest 3B HGRN2 model slightly outperforms Mamba and LLaMa Architecture Transformer for language modeling in a controlled experiment setting; and performs competitively with many open-source 3B models in downstream evaluation while using much fewer total training tokens.