---
id: 20261028-T0-18
title: "TraceQuery：从完整去噪轨迹学习哪些token该重算，降低扩散LLM推理成本"
title_en: "TraceQuery Learns Which Tokens to Recompute in Diffusion LLMs"
url: https://ai.daily.yangsir.net/daily/20261028-T0-18
issue_date: 2026-10-28
publish_date: 2026-10-27T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2610.33008
---

# TraceQuery：从完整去噪轨迹学习哪些token该重算，降低扩散LLM推理成本

arXiv新论文针对扩散大语言模型（dLLM）推理成本高的问题提出TraceQuery。dLLM通过迭代掩码去噪生成文本，但每步都要重复计算token表示，开销随去噪步数累积。现有缓存方法倾向于复用计算结果，TraceQuery则从完整去噪轨迹中学习哪些token需要重新计算、哪些可以复用，从而在保持生成质量的同时减少计算量。目前为论文阶段，未公布基准测试的完整对比数据。

## English Version

**TraceQuery Learns Which Tokens to Recompute in Diffusion LLMs**

A new arXiv paper introduces TraceQuery to cut inference cost in diffusion large language models (dLLMs). dLLMs generate text through iterative masked denoising but recompute token representations at each step, accumulating overhead. Where existing cache methods lean on reuse, TraceQuery learns from the full denoising trajectory which tokens need recomputation and which can be reused, reducing compute while preserving generation quality. It remains a paper, with full benchmark comparisons not yet released.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2610.33008)

**详情页**：https://ai.daily.yangsir.net/daily/20261028-T0-18

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*