AdaptVision introduces an efficient VLM paradigm that autonomously determines the minimum number of visual tokens required for each sample by employing a coarse-to-fine visual acquisition strategy, leading to superior performance with significantly reduced computational overhead.