arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2601.04766cs.CL
收藏

从第一原理重新审视裁判解码:通过无训练分布分歧方法

JuDi: Revisiting Judge Decoding from First Principles via Training-Free Distributional Divergence

City University of Hong Kong(香港城市大学);Huawei Technologies(华为技术有限公司)

Shengyin Sun, Yiming Li, Renxi Liu, Weizhe Lin, Hui-Ling Zhen, Xianzhi Yu, Mingxuan Yuan, Chen Ma

AI总结:

本研究提出一种基于KL散度的无训练裁判解码方法,通过理论证明与实验验证,展示了其在提升推理和编码任务性能上的优势。

摘要:

裁判解码通过放松推测解码的严格验证来加速大语言模型推理,但通常依赖于昂贵且嘈杂的监督。在本工作中,我们从第一原理重新审视这一范式,揭示通过昂贵监督学习的“关键性”分数本质上编码在草稿-目标分布分歧中。我们理论证明了学习的线性裁判与KL散度之间存在结构对应关系,证明它们依赖于相同的底层logit原始数据。基于此,我们提出了一种基于KL散度的简单无训练验证机制。在推理和编码基准上的广泛实验表明,我们的方法在匹配或超越复杂训练裁判(如AutoJudge)的同时,表现出对领域偏移的更强鲁棒性,并完全消除了监督瓶颈。

英文摘要:

Judge Decoding accelerates LLM inference by relaxing the strict verification of Speculative Decoding, yet it typically relies on expensive and noisy supervision. In this work, we revisit this paradigm from first principles, revealing that the ``criticality'' scores learned via costly supervision are intrinsically encoded in the draft-target distributional divergence. We theoretically prove a structural correspondence between learned linear judges and Kullback-Leibler (KL) divergence, demonstrating they rely on the same underlying logit primitives. Guided by this, we propose a simple, training-free verification mechanism based on KL divergence. Extensive experiments across reasoning and coding benchmarks show that our method matches or outperforms complex trained judges (e.g., AutoJudge), offering superior robustness to domain shifts and eliminating the supervision bottleneck entirely. Code is available at https://github.com/sunshy-1/JuDi

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。

↑