SiGMA: Sign-Guided Merging and Adaptation for Multimodal Continual Instruction Tuning

This paper introduces SiGMA, a framework for multimodal continual instruction tuning that combines sign-guided adaptive tuning during training with sign-guided parameter merging at inference. The method aims to reduce collisions between updates for new and previously learned tasks, limiting drift while selectively scaling salient parameters to preserve or amplify task-specific knowledge. The authors report improved performance over existing methods on the UCIT and DCL benchmarks and reduced negative interference. The supplied abstract does not include exact scores, baselines, ablations, model sizes, or computational costs.