Skip to content
Hugging Face Blog

Safety alignment should refuse the harmful subset of a topic, not the whole topic

Safety for Whom? Refusing the Right Subset of a Topic, Not the Whole Topic

Multiverse Computing's new paper argues that current safety alignment treats entire topics as refusal units—LlamaGuard-3, for instance, labels elections as 'factually incorrect information,' causing models to refuse even benign queries. They propose 'narrow-boundary safety': within a single topic like politics, refuse only the harmful subset (e.g., writing targeted manipulation) while still answering benign questions (e.g., election facts). The method uses deployment-specific boundary labels to self-distill a model that respects per-setting splits instead of topic-level blocks. Experiments focus on politics; the post doesn't disclose generalization results for other topics.

Why it matters: A boundary-aware self-distillation method lets a model refuse only the harmful subset of a topic that clashes with deployment policy, not the whole topic.

Read the original ↗Export Markdown