Haokun Lin

Outlier-aware rotation can halve the rotation cost in quantization while maintaining state-of-the-art performance in LLMs.

Existing image editing models fall short when it comes to precise spatial manipulations, but a new benchmark and dataset reveal the path to closing the gap.

VLA models can now be efficiently quantized *without* retraining, even surpassing full-precision performance, thanks to a new post-training method that carefully calibrates scales across attention and output heads.