Reproduce “Generative Models as a Complex Systems Science: How can we make sense of large language model behavior?” on OpenResearch

About this paper

Ari Holtzman, Peter West, Luke Zettlemoyer

Researchers from the University of Washington propose reframing the study of large generative models, such as LLMs, as a complex systems science, moving beyond an engineering-focused approach. The work formally defines “behavior” and introduces a top-down methodology to systematically discover and categorize model capabilities, emphasizing the critical role of open-source models for rigorous scientific inquiry.

Key insights

• Large generative models exhibit characteristics of complex systems, including emergent behaviors that necessitate a scientific approach beyond traditional engineering and performance metrics.
• Understanding "what models do" (their behaviors) should precede and guide attempts to explain "why they do it" (their internal mechanisms and underlying circuits).
• Open-source models are fundamental for conducting scientific research on generative AI, as proprietary APIs restrict critical avenues for transparency, reproducibility, and targeted experimentation.

Summary by alphaXiv.