Saved in:
Bibliographic Details
Main Authors: Xu, Yang, Wang, Yi, Huang, Hengguan, Wang, Hao
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2412.17626
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912410137264128
author Xu, Yang
Wang, Yi
Huang, Hengguan
Wang, Hao
author_facet Xu, Yang
Wang, Yi
Huang, Hengguan
Wang, Hao
contents Understanding training dynamics and feature evolution is crucial for the mechanistic interpretability of large language models (LLMs). Although sparse autoencoders (SAEs) have been used to identify features within LLMs, a clear picture of how these features evolve during training remains elusive. In this study, we (1) introduce SAE-Track, a novel method for efficiently obtaining a continual series of SAEs, providing the foundation for a mechanistic study that covers (2) the semantic evolution of features, (3) the underlying processes of feature formation, and (4) the directional drift of feature vectors. Our work provides new insights into the dynamics of features in LLMs, enhancing our understanding of training mechanisms and feature evolution. For reproducibility, our code is available at https://github.com/Superposition09m/SAE-Track.
format Preprint
id arxiv_https___arxiv_org_abs_2412_17626
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Tracking the Feature Dynamics in LLM Training: A Mechanistic Study
Xu, Yang
Wang, Yi
Huang, Hengguan
Wang, Hao
Machine Learning
Computation and Language
Understanding training dynamics and feature evolution is crucial for the mechanistic interpretability of large language models (LLMs). Although sparse autoencoders (SAEs) have been used to identify features within LLMs, a clear picture of how these features evolve during training remains elusive. In this study, we (1) introduce SAE-Track, a novel method for efficiently obtaining a continual series of SAEs, providing the foundation for a mechanistic study that covers (2) the semantic evolution of features, (3) the underlying processes of feature formation, and (4) the directional drift of feature vectors. Our work provides new insights into the dynamics of features in LLMs, enhancing our understanding of training mechanisms and feature evolution. For reproducibility, our code is available at https://github.com/Superposition09m/SAE-Track.
title Tracking the Feature Dynamics in LLM Training: A Mechanistic Study
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2412.17626