New method deciphers feature flow in foundation models

Researchers have developed a new method to understand how foundation models evolve through fine-tuning and editing, focusing on the internal computations of sparse autoencoders (SAEs). By constructing a transition atlas of feature interactions, they identified thousands of strong ablation-effect transitions in Pythia-160M and Gemma-3-4B models. A significant portion of these transitions showed low cosine similarity between state-target and update-target features, suggesting that standard similarity metrics may not fully capture feature flow dynamics. The findings indicate that feature flow atlases can serve as diagnostics for steering model updates. AI

IMPACT Provides new diagnostic tools for understanding and steering model updates, potentially improving interpretability and control.

RANK_REASON Academic paper detailing a new method for analyzing internal model computations. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gemma 3-4B
• Gotit.pub
• Hugging Face
• multilayer perceptron
• Pythia-160M
• ScienceCast

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →