Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuning

TL;DR

This paper introduces CAFT, a method that uses interpretability tools to ablate undesired concepts in large language models during fine-tuning, effectively steering their out-of-distribution generalization without changing training data.

Contribution

CAFT is a novel technique that leverages concept ablation via linear projections to control LLM generalization during fine-tuning without data modification.

Findings

Reduces misaligned responses by 10x in emergent misalignment scenarios

Maintains performance on training distribution

Does not require modification of training data

Abstract

Fine-tuning large language models (LLMs) can lead to unintended out-of-distribution generalization. Standard approaches to this problem rely on modifying training data, for example by adding data that better specify the intended generalization. However, this is not always practical. We introduce Concept Ablation Fine-Tuning (CAFT), a technique that leverages interpretability tools to control how LLMs generalize from fine-tuning, without needing to modify the training data or otherwise use data from the target distribution. Given a set of directions in an LLM's latent space corresponding to undesired concepts, CAFT works by ablating these concepts with linear projections during fine-tuning, steering the model away from unintended generalizations. We successfully apply CAFT to three fine-tuning tasks, including emergent misalignment, a phenomenon where LLMs fine-tuned on a narrow task generalize to give egregiously misaligned responses to general questions. Without any changes to the fine-tuning data, CAFT reduces misaligned responses by 10x without degrading performance on the training distribution. Overall, CAFT represents a novel approach for steering LLM generalization without modifying training data.