Recycling a target model's own predictions offers a training-free source of proposals for speculative decoding. In vision-language models, however, existing …
机构:中科院
来源:arXiv 2611.05894 | AI4Papers 论文推荐平台