arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2610.29013cs.CLcs.LG
收藏

目标再生会缩小推测解码中草稿训练目标的增益

Target Regeneration Shrinks the Gains of Draft-Training Objectives in Speculative Decoding

Korea University(高丽大学);Yonsei University Mirae Campus(延世大学未来校区)

Jungseob Lee, Hogun Seong, Euijun Choi, Sugyeong Eo

AI总结:

该研究探究推测解码中草稿训练目标的增益因目标再生而缩小的现象,通过多组实验验证该缩小的普遍性,提出需针对目标再生基线并跨训练长度测量目标增益。

摘要:

推测解码通过并行验证草稿 token 来加速大语言模型。新的草稿训练目标通常因在外部编写的响应上比交叉熵(cross-entropy)增加的接受率而受到认可,但许多流水线会让目标重写这些草稿。我们探究这种增益能保留多少。以一个视觉-语言目标和一个块草稿器为基础,我们在一个语料库的四个版本上训练了三种目标:语料库 token 上的交叉熵、从目标进行的 token 级蒸馏,以及对能通过验证的位置进行门控的蒸馏。蒸馏和门控在外部响应上带来了大幅增益,但在贪婪再生后,蒸馏的增益变小,门控几乎没有增益。这种缩小现象在数据顺序、位置权重、验证器、解码温度,以及对于蒸馏而言的初始化和学习率上均存在。在匹配的监督长度和各目标的最佳测量训练长度下,大部分缩小现象仍然存在。第二个视觉语料库和仅文本的复现实验也显示了这种缩小现象,对于蒸馏而言还使用了自回归草稿器。因此,目标增益应针对目标再生基线并在不同训练长度下进行测量。我们的代码可在该 https URL 获取。

英文摘要:

Speculative decoding accelerates large language models by verifying drafted tokens in parallel. New draft-training objectives are usually credited with the acceptance they add over cross-entropy on externally written responses, yet many pipelines let the target rewrite them. We ask how much of that gain survives. With a vision-language target and a block drafter, we train three objectives on four versions of one corpus: cross-entropy on corpus tokens, token-level distillation from the target, and distillation gated to positions that would survive verification. Distillation and the gate add large gains on external responses, but after greedy regeneration, distillation adds a smaller gain and the gate little or none. The shrinkage persists across data orders, position weights, verifiers, decoding temperatures and, for distillation, initializations and learning rates. Most of it remains at matched supervised length and at each objective's best measured training length. A second vision corpus shows the shrinkage, as does a text-only replication, for distillation also with an autoregressive drafter. Objective gains should therefore be measured against a target-regenerated baseline and across training lengths. Our code is available at https://github.com/js-lee-AI/TargetRegeneration.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑