Daily Papers

Safety fine-tuning algorithms are commonly used to fine-tune language models to reduce harmful outputs, but the exact internal mechanisms of how those models achieve this remain unclear. In studying direct preference optimisation (DPO) for toxicity reduction, current explanations claim that DPO works by dampening the most toxic MLP neurons to learn an offset to avert toxic regions in the residual stream. However, by ablating the most toxic neurons and applying activation patching, we find this explanation incomplete. By projecting neuron activation changes onto a toxicity probe, we find that only 31.8\% of toxicity reduction comes from dampened toxic neurons. Instead, DPO reduces toxicity by accumulating effects across multiple neuron groups, both reducing writing in the toxic direction and promoting anti-toxicity in the residual stream. Moreover, DPO gives noisy adjustments to neuron activations, with many neurons actually increasing toxicity. This indicates that DPO is a balancing process between opposing neuron effects to achieve toxicity reduction.

Text detoxification has the potential to mitigate the harms of toxicity by rephrasing text to remove offensive meaning, but subtle toxicity remains challenging to tackle. We introduce MaRCo, a detoxification algorithm that combines controllable generation and text rewriting methods using a Product of Experts with autoencoder language models (LMs). MaRCo uses likelihoods under a non-toxic LM (expert) and a toxic LM (anti-expert) to find candidate words to mask and potentially replace. We evaluate our method on several subtle toxicity and microaggressions datasets, and show that it not only outperforms baselines on automatic metrics, but MaRCo's rewrites are preferred 2.1 times more in human evaluation. Its applicability to instances of subtle toxicity is especially promising, demonstrating a path forward for addressing increasingly elusive online hate.