A preprint reports a 13.7-point shallow-layer POPE F1 gap and model-specific latency results for visual-token pruning in AI models.