Constructive Distortion: Improving MLLMs with Attention-Guided Image Warping
The episode discusses a paper titled "Constructive Distortion: Improving MLLMs with Attention-Guided Image Warping." Hosts discuss how this method uses an MLLM's attention signals to dynamically warp input images at test time, expanding high-attention regions. This technique improves fine-grained perception and spatial reasoning without retraining the model, showing strong performance gains over existing methods.