Tag
This paper proposes drift-constrained optimization for fine-tuning instruct models, focusing on the direction of updates to improve target-task performance while preserving existing capabilities. Experiments show that layer-selective probing can enhance scientific reasoning and multilingual translation across multiple languages.