SGM: "Safety glasses" for multimodal large language models via neuron-level detoxification

SGM is a white-box neuron-level detoxification intervention for multimodal large language models: it recalibrates toxicity-related neurons using expert-weighted soft suppression, neutralizing harmful cross-modal activations without updating parameters. The study establishes the MM-TOXIC-QA framework and reduces the average harmfulness rate from 45.0% to 4.5% under standard and adversarial conditions on open-source MLLMs, while maintaining fluency and multimodal reasoning. SGM* can be combined with existing detoxification methods to enhance safety performance.