Xiangyu Zhou

LVLMs can now provide depth-aware pedestrian navigation guidance by grounding language reasoning and segmentation, without needing user-provided cues or anchor points.

Forget sensitive data without sacrificing model coherence: Attention Smoothing Unlearning (ASU) leverages self-distillation with temperature-scaled attention to erase unwanted knowledge while preserving general language abilities.

Standard image segmentation techniques undermine the theoretical guarantees of hierarchical Shapley explanations, but a new segmentation method restores consistency and boosts performance.