M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling

TL;DR

This paper introduces M$^2$RNN, a matrix-valued non-linear RNN architecture that enhances language modeling by enabling better state tracking and scalability, outperforming hybrid models with fewer resources.

Contribution

The paper proposes M$^2$RNN with matrix-valued states and state size expansion, demonstrating improved performance and generalization in large-scale language modeling tasks.

Findings

M$^2$RNN achieves perfect state tracking at unseen sequence lengths.

Hybrid M$^2$RNN outperforms Gated DeltaNet by 0.4-0.5 perplexity points on 7B MoE.

Replacing a single recurrent layer with M$^2$RNN improves accuracy with minimal throughput impact.

Abstract

Transformers are highly parallel but are limited to computations in the TC$^0$ complexity class, excluding tasks such as entity tracking and code execution that provably require greater expressive power. Motivated by this limitation, we revisit non-linear Recurrent Neural Networks (RNNs) for language modeling and introduce Matrix-to-Matrix RNN (M$^2$RNN): an architecture with matrix-valued hidden states and expressive non-linear state transitions. We demonstrate that the language modeling performance of non-linear RNNs is limited by their state size, and show how the state size expansion mechanism enables efficient use of tensor cores. Empirically, M$^2$RNN achieves perfect state tracking generalization at sequence lengths not seen during training. These benefits also translate to large-scale language modeling. In hybrid settings that interleave recurrent layers with attention, Hybrid M$^2$RNN outperforms equivalent Gated DeltaNet hybrids by $0.4$-$0.5$ perplexity points on a 7B MoE model, while using $3\times$ smaller state sizes for the recurrent layers. Notably, replacing even a single recurrent layer with M$^2$RNN in an existing hybrid architecture yields accuracy gains comparable to Hybrid M$^2$RNN with minimal impact on training throughput. Further, the Hybrid Gated DeltaNet models with a single M$^2$RNN layer also achieve superior long-context generalization, outperforming state-of-the-art hybrid linear attention architectures by up to $8$ points on LongBench. Together, these results establish non-linear RNN layers as a compelling building block for efficient and scalable language models.