Tag
The paper introduces GLANCE, a one-pass block drafting method for lossless speculative decoding in vision-language models, achieving up to 2.93x faster generation without changing output.