A Mechanistic Investigation of Supervised Fine Tuning

TL;DR

This paper introduces a new diagnostic pipeline using pretrained Sparse Autoencoders to analyze how supervised fine-tuning alters semantic features in language models, revealing task-specific and safety alignment layer effects.

Contribution

It presents a novel SAE-based method to mechanistically investigate the representational changes caused by supervised fine-tuning in language models.

Findings

SAE projections reveal significant divergence in sparse latents post-fine-tuning.

Identifies task-specific and layer-specific semantic feature alterations.

Discovers a layer-wise update profile associated with safety alignment.

Abstract

The cosine similarity between a large language model's hidden activations before and after Supervised Fine-Tuning (SFT) remains very high. This, at first glance, suggests that SFT leaves the model's activation geometry largely undisturbed. However, projecting both sets of activations through a Sparse Autoencoder (SAE) pretrained on the base model reveals that the underlying sparse latents diverge significantly. We introduce a novel investigative pipeline which utilizes these pretrained SAEs as a high-resolution diagnostic tool to mechanistically investigate the drivers of this representational divergence. Through our analytical pipeline, we discover task-specific and layer-specific distributions of the precise semantic features that are systematically altered during supervised fine-tuning. We additionally identify a layer-wise update profile specific to safety alignment. All code, experimental scripts, and analysis files associated with this work are publicly available at: https://github.com/ruhzi/sae-investigation.