1
SIEVES: Selective Prediction Generalizes through Visual Evidence Scoring
选择性预测遇上视觉证据打分！SIEVES让多模态大模型在零样本领域外基准上覆盖率提升3倍，且无需模型内部信号，闭源模型也能直接用——这才是可靠部署的真正解法。
arXiv:2604.25855v2 Announce Type: replace-cross Abstract: Multimodal large language models (MLLMs) achieve ever-stronger performance on visual-languag…