ISSN :2582-9793
  • Subscribe

    • Going Beyond a Basic Attention Head toward an Understanding of Transformer-based Generative AI

    Original Research (Published On: 14-Dec-2025 )
    DOI : https://doi.org/10.54364/AAIML.2025.54259

    Neil Johnson, Nicholas Restrepo, Frank Huo and Dylan Restrepo

    Adv. Artif. Intell. Mach. Learn., 5 (4):4675-4691

    1. Neil Johnson: George Washington University

    2. Nicholas Restrepo: George Washington University

    3. Frank Huo: George Washington University

    4. Dylan Restrepo: Cornell University

    Download PDF Here

    DOI: 10.54364/AAIML.2025.54259

    Article History: Received on: 24-Aug-25, Accepted on: 07-Dec-25, Published on: 14-Dec-25

    Corresponding Author: Neil Johnson

    Email: neilfjohnson@me.com

    Citation: Nicholas J. Restrepo, et al. Going Beyond a Basic Attention Head toward an Understanding of Transformerbased Generative AI. Advances in Artificial Intelligence and Machine Learning. 2025. (Ahead of print). https://dx.doi.org/10.54364/AAIML.2025.54259


    Abstract

        

    This paper goes beyond the basic Attention head analysis introduced in our accompanying paper, as part of our long-term goal to establish a bottom-up understanding of Chat-GPT-like generative AI. We provide evidence that suggests the output tipping behavior that we reported for the basic Attention head can persist despite some of the complications of real LLMs. Specifically, we consider here (1) a richer vocabulary, (2) non-identity matrices learned during pre-training, (3) non-zero temperature during next-token selection, (4) more than one layer of Attention heads. We then offer some preliminary evidence that the insights gained from this bottom-up approach, can help improve performance in real-world generative AI systems.


    Statistics

       Article View: 1158
       PDF Downloaded: 10