Top-Down Compression: Revisit Efficient Vision Token Projection for Visual Instruction Tuning Paper • 2505.11945 • Published May 17 • 5
