Circuit Breaking: Removing Model Behaviors with Targeted Ablation

TL;DR

This paper introduces a targeted ablation method to remove undesirable behaviors in language models by disabling specific causal pathways, effectively reducing toxic outputs with minimal impact on overall performance.

Contribution

It presents a novel approach to identify and ablate a small set of causal pathways to mitigate harmful behaviors in language models.

Findings

Ablating 12 causal edges reduces GPT-2 toxic language generation

Minimal performance degradation on other inputs

Effective removal of undesirable behaviors with small ablation

Abstract

Language models often exhibit behaviors that improve performance on a pre-training objective but harm performance on downstream tasks. We propose a novel approach to removing undesirable behaviors by ablating a small number of causal pathways between model components, with the intention of disabling the computational circuit responsible for the bad behavior. Given a small dataset of inputs where the model behaves poorly, we learn to ablate a small number of important causal pathways. In the setting of reducing GPT-2 toxic language generation, we find ablating just 12 of the 11.6K causal edges mitigates toxic generation with minimal degradation of performance on other inputs.