Reproduce “Fine-Tuning is Subgraph Search: A New Lens on Learning Dynamics” on OpenResearch

About this paper

Yueyan Li, Wenhao Gao, Caixia Yuan, Xiaojie Wang

Researchers from Beijing University of Posts and Telecommunications propose "circuit-tuning," a framework that redefines fine-tuning as a dynamic search for task-specific computational subgraphs within large language models. This method dynamically identifies and updates only relevant model components, achieving strong task performance while better preserving general capabilities across various language tasks.

Key insights

• Fine-tuning can be conceptualized as a "subgraph search," where only a small, task-specific computational circuit is intrinsically necessary for learning a new ability.
• The learning process involves dynamic adjustment of functional roles within model components, with some heads functionally reversing while others maintain their core identification roles.
• Focusing parameter updates on an identified "intrinsic graph" helps balance target task performance with the preservation of general model capabilities.

Summary by alphaXiv.