Skip to yearly menu bar Skip to main content


Mechanistic Design and Scaling of Hybrid Architectures

Michael Poli · Armin Thomas · Eric Nguyen · Pragaash Ponnusamy · Bj√∂rn Deiseroth · Kristian Kersting · Taiji Suzuki · Brian Hie · Stefano Ermon · Christopher Re · Ce Zhang · Stefano Massaroli

Hall C 4-9
[ ]
Tue 23 Jul 4:30 a.m. PDT — 6 a.m. PDT


The development of deep learning architectures is a resource-demanding process, due to a vast design space, long prototyping times, and high compute costs associated with at-scale model training and evaluation. We set out to simplify this process by grounding it in an end-to-end mechanistic architecture design (MAD) pipeline, encompassing small-scale capability unit tests predictive of scaling laws. Through a suite of synthetic token manipulation tasks such as compression and recall, designed to probe capabilities, we identify and test new hybrid architectures constructed from a variety of computational primitives. We experimentally validate the resulting architectures via an extensive compute-optimal and a new state-optimal scaling law analysis, training over 500 language models between 70M to 7B parameters. Surprisingly, we find MAD synthetics to correlate with compute-optimal perplexity, enabling accurate evaluation of new architectures via isolated proxy tasks. The new architectures found via MAD, based on simple ideas such as hybridization and sparsity, outperform state-of-the-art Transformer, convolutional, and recurrent architectures (Transformer++, Hyena, Mamba) in scaling, both at compute-optimal budgets and in overtrained regimes. Overall, these results provide evidence that performance on curated synthetic tasks can be predictive of scaling laws, and that an optimal architecture should leverage specialized layers via a hybrid topology.

Live content is unavailable. Log in and register to view live content