MOSAIC: Composable Safety Alignment via Learnable Control Tokens Over Frozen LLM Backbone

arXiv 2603.16210·medium signal

MOSAIC enables context-dependent safety rules that vary by user, region, or application — a direct fix for the limitation where static parameter-level alignment entangles safety with capabilities. It trains modular control tokens (each representing one safety constraint) optimized over a frozen backbone, composable at inference time via order-based task sampling. Experiments show strong defense with substantially lower over-refusal rates compared to standard alignment approaches, preserving model utility.