Multi-Modal In-Context Learning (MM-ICL) is commonly employed to enhance performance of Large Vision-Language Models (LVLMs) on downstream tasks. However, as next-generation models like QwenVL and InternVL have been optimized through instruction tuning and alignment, their baseline capabilities have been strengthened. This necessitates a re-evaluation of the effectiveness and optimal configuration of MM-ICL across different models. By evaluating diverse LVLMs spanning different parameter scales (7B–32B) and accessibility settings (open-source and closed-source) across eight cross-domain datasets, this paper systematically investigates the performance of MM-ICL under varying model capabilities and task settings. Experimental results indicate that, beyond validating established patterns, next-generation models exhibit distinct phenomena: in certain logical reasoning and format-sensitive tasks, the introduction of MM-ICL does not necessarily yield improvements and can even lead to performance degradation in specific scenarios. We analyze these observations and, based on extensive experimental data, provide new recommendations for selecting MM-ICL strategies for LVLMs.