multimodalfactbullishReMo can remove 54% of input tokens in Omni-modal LLMs with no loss in accuracy by redistributing visual information across modalitiesComputer Vision28 Jul 2026http://arxiv.org/abs/2607.21179v1