Safety alignment usually costs utility because the aligned weights apply to every prompt;

New Stanford univ paper finds that scaling the safety update per prompt recovers much of what global tuning gives up.

The problem is that a safety fine-tune changes the model for every input. Harmful or not, every prompt now runs through a safer but weaker model.

CLEAR, proposed in this paper, leaves the original model frozen. A small gate reads each incoming prompt and decides how much of a separate safety module to switch on.

Benign prompts get almost none of it, so they run on the untouched model.

– arxiv. org/abs/2608.21278

Title: "CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment"