The Transformer Revolution, Part 1: Dynamic Process... | AI Research

The paper "The Transformer Revolution, Part 1: Dynamic Processing through Output- Weight Interconnections" by Marco Giunti and Fabrizia Giulia Garavaglia proposes that Transformers function by generating prompt-dependent transformations during inference, rather than simply reproducing statistical patterns learned during training. The authors introduce a framework called Sequence-level Interactive Dynamic Parallel Processing (SIDPP) to describe this mechanism.

Redefining Transformer Computation

The authors challenge the "stochastic parrot" view of large language models, which suggests these systems merely repeat learned statistical regularities. Instead, they argue that Transformers act as systems that transform concepts using other concepts. In this model, token vectors represent the concepts being transformed, while parameterized transformations—defined by matrices and vectors—act as the transforming concepts. These transformations can be static (fixed during training) or dynamic (generated from the input sequence).

The Role of Output-Weight Interconnections

The core architectural innovation identified by the authors is the use of "output-weight interconnections." In standard neural networks, outputs typically flow to the inputs of subsequent layers. In this proposed model, the outputs of certain networks also determine the weights of others. By using these interconnections, the Transformer constructs specific transformations based on the input prompt, which are then applied to modify token representations.

Strong Prompt Sensitivity

The authors introduce the concept of "strong prompt sensitivity" to describe how the contribution of dynamic processing changes based on the input. They observe that the influence of dynamic processing grows as the prompt length increases. According to the paper, this dynamic contribution can eventually equal or exceed the contribution of static processing.

Implications for Future Research

The authors suggest that this interpretation of SIDPP has practical implications for improving the interpretability, predictability, and control of language models. Furthermore, they propose that this framework could assist in the design of smaller, more sustainable systems. Finally, the authors conjecture that because the human neural system possesses the necessary mechanisms for SIDPP, human language processing might itself be a form of SIDPP, potentially sharing a functional architecture similar to that of the Transformer.