arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2510.05040cs.LGcs.AI
收藏

扩散语言模型中的测试时扩展:基于隐藏半自回归专家

Test-Time Scaling in Diffusion LLMs via Hidden Semi-Autoregressive Experts

Yonsei University(延世大学);Oracle(Oracle公司);CMU(卡内基梅隆大学);UMD(马里兰大学);UCF(佛罗里达大学)

Jihoon Lee, Hoyeon Moon, Kevin Zhai, Arun Kumar Chithanar, Anit Kumar Sahu, Soummya Kar, Chul Lee, Souradip Chakraborty, Amrit Singh Bedi

AI总结:

提出HEX方法,通过集成不同块大小的生成路径进行多数投票,实现无需训练的测试时扩展,显著提升扩散语言模型在推理基准上的性能。

摘要:

基于扩散的大语言模型(dLLMs)经过灵活训练,以建模数据分布中的极端依赖性;然而,在推理时如何最好地利用这些信息仍然是一个开放问题。在这项工作中,我们揭示了这些模型的一个有趣特性:在文本数据上训练的dLLMs隐式地学习了一个半自回归专家的混合体,其中不同的生成顺序揭示了不同的专门化行为。我们表明,采用任何单一的、固定的推理时间调度(一种常见做法)会因未能利用这一潜在集成而导致性能崩溃。为了解决这个问题,我们引入了HEX(用于测试时扩展的隐藏半自回归专家),这是一种无需训练的推理方法,它跨异构块调度进行集成。通过对不同块大小的生成路径进行多数投票,HEX稳健地避免了与任何单一固定调度相关的失败模式。在诸如GSM8K等推理基准上,它将准确率提升了高达3.56倍(从24.72%提升至88.10%),优于top-K边际推理和专门微调方法(如GRPO),且无需额外训练。HEX还在MATH基准上取得了显著提升,从16.40%提升至40.00%;在ARC-C科学推理上从54.18%提升至87.80%;在TruthfulQA上从28.36%提升至57.46%。我们的结果为基于扩散的语言模型(dLLMs)中的测试时扩展建立了新的范式,揭示了掩码执行的顺序在决定推理性能中起着关键作用。

英文摘要:

Diffusion-based large language models (dLLMs) are trained flexibly to model extreme dependence in the data distribution; however, how to best utilize this information at inference time remains an open problem. In this work, we uncover an interesting property of these models: dLLMs trained on textual data implicitly learn a mixture of semi-autoregressive experts, where different generation orders reveal different specialized behaviors. We show that committing to any single, fixed inference time schedule, a common practice, collapses performance by failing to leverage this latent ensemble. To address this, we introduce HEX (Hidden semiautoregressive EXperts for test-time scaling), a training-free inference method that ensembles across heterogeneous block schedules. By doing a majority vote over diverse block-sized generation paths, HEX robustly avoids failure modes associated with any single fixed schedule. On reasoning benchmarks such as GSM8K, it boosts accuracy by up to 3.56X (from 24.72% to 88.10%), outperforming top-K margin inference and specialized fine-tuned methods like GRPO, without additional training. HEX even yields significant gains on MATH benchmark from 16.40% to 40.00%, scientific reasoning on ARC-C from 54.18% to 87.80%, and TruthfulQA from 28.36% to 57.46%. Our results establish a new paradigm for test-time scaling in diffusion-based LLMs (dLLMs), revealing that the sequence in which masking is performed plays a critical role in determining performance during inference.

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。

↑