multimodalfactneutralLong visual context causes performance degradation in vision-language models as the number of distractors growsMachine Learning02 Aug 2026http://arxiv.org/abs/2607.28627v1