Training-Free Loosely Speculative Decoding: Accepting Semantically Correct Drafts Beyond Exact Match
无需训练的宽松推测解码:在精确匹配之外接受语义正确的草稿
机构 * Advanced Micro Devices, Inc.(先进微器件公司) ; Department of Electrical and Electronic Engineering, The University of Hong Kong(香港大学电子与电气工程系)
AI总结 本文提出无需训练的宽松推测解码方法FLy,通过利用目标模型的自我纠正行为来判断草稿-目标不匹配是否仍具有语义有效性,从而在保持高精度的同时提升解码效率。
Comments Published as a conference paper at ICLR 2026