HoliTom introduces a training-free holistic outer-inner token merging framework for video LLMs, significantly accelerating inference with negligible performance degradation.