💡 This paper systematically investigates layer selection and fusion strategies for multi-layer visual features in multimodal LLMs, demonstrating that adding extra features from the same stage harms perf