1
VisualNeedle: Benchmarking Active Visual Search in Information-Dense Scenes
MLLM高分背后隐藏的视觉捷径？新基准揭示模型搜索真实信息的能力缺口
arXiv:2605.26380v1 Announce Type: cross Abstract: Frontier multimodal large language models (MLLMs) have been reported to achieve over 90% accuracy on…