coco-22c59c84·2 events·first seen Aliases: COCO
COCO
MixFrag is a new post-training quantization (PTQ) framework for Vision Transformers that assigns mixed bit-widths per layer based on quantization fragility, measured via KL divergence between full-precision and quantized output distributions. Bit allocation is formulated as a Multiple-Choice Knapsack Problem to optimize precision under a target bit budget. Evaluated on ImageNet-1K classification and COCO detection/segmentation, MixFrag claims state-of-the-art among mixed-precision PTQ methods, improving the prior best by up to 9.6 AP on a challenging low-bit setting.
Researchers identify a small set of attention heads in vision-language model backbones, called 'gaze heads', whose attention patterns track the image region currently being described. Using comic strips as a controlled testbed, they show that intervening on the top-100 gaze heads (fewer than 9% of all heads) can steer the model to describe any chosen region at 83.1% accuracy, without retraining. The mechanism generalizes across model sizes from 2B to 32B parameters and to natural images (COCO), establishing a practical inference-time control lever for multimodal models via mechanistic analysis.