On the Need to Compare the Architecture Design of Large Language Models: A Preliminary Reference Metamodel

• Jesús Carreño-Bolufer Universitat Politècnica de València
• Giovanni Giachetti Universitat Politècnica de València / Universidad Andrés Bello

Resumo

A large number of Large Language Models (LLMs) have been developed and made available to the community. However, analyzing and comparing these models remains challenging since there is no clear, standardized way to describe their architectures precisely. While LLM developers typically provide textual model cards, configuration files, or source code, these artifacts do not follow a common standard. Consequently, researchers need to make additional efforts to reuse and compare existing LLMs. To address this issue, we propose a conceptual modeling approach to create a model-driven method for describing and analyzing LLM architectures.

Referências

Minaee, S. et al. (2024). Large language models: A survey.

Object Management Group (2016). Meta-Object Facility (MOF). Object Management Group, 2.5.1 edition. Formal specification.

Pastor, O., España, S., Panach, J. I., and Aquino, N. (2008). Model-driven development. Informatik-Spektrum, 31(5):394–407.

Vaswani, A. et al. (2017). Attention is all you need.

Wieringa, R. (2014). Design science methodology for information systems and software engineering. Springer.