Beyond Surrogate Gradients: Fully Differentiable Token Pruning for Vision-Language Models 文章

ArXiv CS.CV2026-05-28NEWSen作者: Landi He, Mingde Yao, Shawn Young, Lijian Xu
查看原文 →
关系图谱

Beyond Surrogate Gradients: Fully Differentiable Token Pruning for Vision-Language Models · 相关技术

相关技术

DiffPruneVision-Language ModelsVisual token pruningInformation ThrottlerGumbel-Softmax