Transformer Layers as Painters

TL;DR

This paper investigates the internal structure of transformer models, revealing layer-specific roles and robustness to modifications, which could inform more efficient model usage and design.

Contribution

It provides empirical insights into layer functions and robustness in frozen transformers, suggesting potential for latency reduction without significant accuracy loss.

Findings

Lower and final layers differ from middle layers

Middle layers exhibit uniformity across models

Models can maintain performance when skipping or reordering layers

Abstract

Despite their nearly universal adoption for large language models, the internal workings of transformers are not well understood. We aim to better understand the impact of removing or reorganizing information throughout the layers of a pretrained transformer. Such an understanding could both yield better usage of existing models as well as to make architectural improvements to produce new variants. We present a series of empirical studies on frozen models that show that the lower and final layers of pretrained transformers differ from middle layers, but that middle layers have a surprising amount of uniformity. We further show that some classes of problems have robustness to skipping layers, running the layers in an order different from how they were trained, or running the layers in parallel. Our observations suggest that even frozen pretrained models may gracefully trade accuracy for latency by skipping layers or running layers in parallel.