Precisely Characterizing Fine-Tuning Generalization via Diagonal Linear Networks

ClementineDomi6 · x · 2026-07-08

This study uses diagonal linear networks as an analytically tractable model to precisely characterize fine-tuning behavior. By employing replica theory, it derives exact generalization curves, linking initialization, data structure, and sample efficiency.

Core findings: Pretraining hyperparameters shape the model's inductive bias, determining whether fine-tuning reuses pretrained features or learns new ones. The relative scale of weights across layers acts as a key control knob. When a task relies on a sparse subset of pretrained features, a negative relative scale can improve generalization. This conclusion extends to ResNets and Transformers in non-ideal pretraining settings.

From this, the study derives four distinct fine-tuning mechanisms, including a previously overlooked one that simultaneously achieves both reuse and refinement.

Related event: ICML Paper: How Pretraining Shapes Inductive Bias in Fine-Tuning(5 posts)→

Original post →

More from Research

Research channel →