Circuit Breaking: Removing Model Behaviors with Targeted Ablation
This addresses the issue of harmful model outputs for users of language models, offering a targeted solution to mitigate specific bad behaviors.
The paper tackles the problem of undesirable behaviors in language models, such as toxic language generation, by proposing a targeted ablation approach that removes specific causal pathways; it demonstrates that ablating just 12 out of 11,600 edges in GPT-2 reduces toxicity with minimal performance degradation.
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.