Implicit Language Models are RNNs: Balancing Parallelization and Expressivity
This work addresses a fundamental bottleneck in language modeling by balancing computational efficiency and model expressivity, offering a novel approach that could impact large-scale AI systems.
The paper tackles the trade-off between parallelization and expressivity in language models by proposing implicit state-space models (SSMs) that approximate RNN-like non-linear transitions, enabling scalable training with superior state-tracking on regular languages and outperforming explicit models on benchmarks, including scaling to 1.3B parameters on 207B tokens.
State-space models (SSMs) and transformers dominate the language modeling landscape. However, they are constrained to a lower computational complexity than classical recurrent neural networks (RNNs), limiting their expressivity. In contrast, RNNs lack parallelization during training, raising fundamental questions about the trade off between parallelization and expressivity. We propose implicit SSMs, which iterate a transformation until convergence to a fixed point. Theoretically, we show that implicit SSMs implement the non-linear state-transitions of RNNs. Empirically, we find that only approximate fixed-point convergence suffices, enabling the design of a scalable training curriculum that largely retains parallelization, with full convergence required only for a small subset of tokens. Our approach demonstrates superior state-tracking capabilities on regular languages, surpassing transformers and SSMs. We further scale implicit SSMs to natural language reasoning tasks and pretraining of large-scale language models up to 1.3B parameters on 207B tokens representing, to our knowledge, the largest implicit model trained to date. Notably, our implicit models outperform their explicit counterparts on standard benchmarks. Our code is publicly available at http://github.com/microsoft/implicit_languagemodels .