论文标题:
Follow the Latent Roadmap: Navigating Revocable Decoding for Diffusion LLMs with Anchor Tokens
发表日期:
2026年06月 发表单位:
King's College London, The Chinese University of Hong Kong, The Alan Turing Institute 原文链接:
https://arxiv.org/pdf/2606.16847v1.pdf 开源代码链接:
未提供 项目链接:
未提供 开源数据集链接:
未提供
验证待定token时,ASRD采用锚点扰动验证:提取 ct 中与待定token嵌入 eti 正交的部分 dj,乘以小系数 β 后加到 eti 上。若token正确,其方向与锚点接近,正交残差小,干扰弱;若基于“局部错误共识”产生错误,方向偏离锚点,正交残差大,打破“局部共识”,导致预测翻转,被标记为失败。这招“打草惊蛇”,精准揪出隐藏错误。
[1] Nie et al., 2025, LLaDA: Large Language Diffusion with Masking.[2] Ye et al., 2025, Dream 7B: A Diffusion Model for Text Generation.[3] Hong et al., 2025, WINO: A Draft-and-Verify Framework for Diffusion LLMs.[4] Dong et al., 2025, Saber: Confidence-Aware Acceleration for Diffusion Models.[5] Cobbe et al., 2021, Training Verifiers to Solve Math Word Problems (GSM8K).[6] Lightman et al., 2024, Let's Verify Step by Step (MATH500).[7] Chen et al., 2021, Evaluating Large Language Models Trained on Code (HumanEval).[8] Austin et al., 2021b, Program Synthesis with Large Language Models (MBPP).[9] Yao et al., 2026, Follow the Latent Roadmap: Navigating Revocable Decoding for Diffusion LLMs with Anchor Tokens. arXiv:2606.16847.