V-GIFT: Boosting Visual Instruction Tuning with Self-Supervised Guidance
V-GIFT improves Multimodal Large Language Models (MLLMs) at solving vision-centric tasks by adding a small amount of visually grounded self-supervised learning (SSL) data to their instruction tuning phase. The method reformulates classical tasks like rotation prediction into image-instruction triplets, forcing the model to use visual evidence instead of relying solely on language shortcuts, leading to consistent performance gains.