Researchers have found that fine-tuning language models for specific tasks can cause them to lose their ability to refuse harmful requests. This is a problem because it makes the models vulnerable to being used in ways that are not safe or intended. A new approach, called RefusalGuard, has been developed to address this issue by preserving the geometric structure of safety-relevant representations within the model's activation space. This allows the model to maintain its refusal behavior while still adapting to the specific task at hand. --- Why it matters: This matters because it can help prevent language models from being used in ways that are not safe or intended, which is a growing concern as these models become more widespread and powerful. By preserving the safety-relevant structure of the model's representations, RefusalGuard can help ensure that language models remain aligned with their intended purpose. Source: https://arxiv.org/abs/2605.01913