Be Confident: Uncovering Overfitting in MLLM Multi-Task Tuning | Read Paper on Bytez