跳转至

文章背景与核心概要

大语言模型（LLMs）在各项推理基准测试中展现出了强大的能力，然而学术界对其在面对现实世界的词汇污染（如键盘误触、字符颠倒和填充词插入等）时的鲁棒性仍缺乏深入理解。本文对多个开源指令微调模型及前沿模型在不同词汇噪声下的表现进行了全面评估，并揭示了字符级扰动如何通过“注意力转移（Attention Diversion）”机制显著降低多步推理的准确率。

研究发现，词汇破坏会打碎子词分词（subword tokenization），导致生成的碎片化分词在Transformer的中层和底层吸引不成比例的注意力权重。通过长度匹配对照实验以及析因干预分析，作者进一步证明了分词碎片化（而非提示词长度）是性能下降的罪魁祸首，且分词内容与注意力分配之间存在深度耦合。这一发现解释了为什么现有的推理时防御手段（如思维链提示、拼写检查等）往往难以奏效，因为它们未能同时修复受损的内容与注意力结构。

Lexical Perturbations Disrupt LLM Reasoning: An Empirical Study of Attention Diversion

Lexical Perturbations Disrupt LLM Reasoning: An Empirical Study of Attention Diversion

license icon

license icon

Summary

Summary

This paper investigates the robustness of Large Language Models (LLMs) against realistic lexical corruptions (such as keyboard noise, character swaps, and filler insertions) across various reasoning benchmarks. The authors demonstrate that character-level perturbations significantly degrade multi-step reasoning accuracy due to a phenomenon termed Attention Diversion: lexical corruption fragments subword tokenization, trapping disproportionate attention mass in middle and final transformer layers.

Length-matched controls prove that token fragmentation—rather than prompt length—drives performance loss. Furthermore, factorial interventions reveal that token content and attention allocation are deeply coupled, explaining why standard inference-time defenses (like chain-of-thought prompting or spell-checking) fail to recover performance on their own.

本文研究了大语言模型（LLMs）在面对各种推理基准测试中现实世界的词汇污染（如键盘误触、字符颠倒和填充词插入）时的鲁棒性。作者证明，字符级扰动会显著降低多步推理的准确率，其背后的核心机制被称为注意力转移（Attention Diversion）：词汇破坏打碎了子词分词，将不成比例的注意力质量困在Transformer的中层和最终层中。

长度匹配对照实验证明，是分词碎片化（而非提示词长度）导致了性能损失。此外，析因干预（factorial interventions）表明，分词内容与注意力分配深度耦合，这解释了为什么标准的推理时防御手段（如思维链提示或拼写检查）无法单独恢复模型性能。

Document Metadata

Document Metadata

Metadata Field Details
arXiv Identifier arXiv:2608.22140 [cs.CL]
Primary Subject Computation and Language (cs.CL)
Secondary Subjects Artificial Intelligence (cs.AI), Machine Learning (cs.LG)
Submission Date August 22, 2026
Authors Jiaqian Zhu, Yang Zhang, Junhua Ding, Xiaowei Yu
Publication Status Accepted to EMNLP 2026 (Main Conference) — 9 pages main text, 12 figures, 20 tables
DOI 10.48550/arXiv.2608.22140
Project Repository GitHub: Attention-Diversion
Metadata Field Details
arXiv Identifier arXiv:2608.22140 [cs.CL]
Primary Subject Computation and Language (cs.CL)
Secondary Subjects Artificial Intelligence (cs.AI), Machine Learning (cs.LG)
Submission Date August 22, 2026
Authors Jiaqian Zhu, Yang Zhang, Junhua Ding, Xiaowei Yu
Publication Status Accepted to EMNLP 2026 (Main Conference) — 9 pages main text, 12 figures, 20 tables
DOI 10.48550/arXiv.2608.22140
Project Repository GitHub: Attention-Diversion

Abstract

Abstract

Large Language Models (LLMs) achieve strong reasoning performance, but their robustness to realistic lexical corruption remains poorly understood. We evaluate four open-weight instruction-tuned models and frontier models across four reasoning benchmarks under keyboard noise, character swaps, and filler insertion. Character-level perturbations substantially degrade accuracy, especially on multi-step reasoning tasks, while filler insertion has little effect. We trace this asymmetry to Attention Diversion: lexical corruption fragments subword tokenization, and the resulting fragments attract disproportionate attention mass, concentrated in middle and final transformer layers. Length-matched controls confirm that fragmentation, not prompt length, drives the loss. A factorial intervention then shows why the damage is hard to undo: fragmentation corrupts token content and attention allocation together, and the two are coupled. Restoring clean attention while the content remains corrupted is actively harmful, restoring content alone is insufficient, and only restoring both recovers a substantial share of the gap. This coupling explains why inference-time strategies, including chain-of-thought prompting, spell-checking, self-repair, and stronger repair models, fail to consistently recover performance: each addresses one channel at a time.

大语言模型（LLMs）具备强大的推理性能，但学术界对其应对现实世界词汇污染的鲁棒性仍缺乏充分理解。我们在四种推理基准测试中，评估了四个开源指令微调模型和前沿模型在键盘噪声、字符交换和填充插入条件下的表现。字符级扰动大幅降低了准确率（特别是在多步推理任务中），而填充插入的影响则微乎其微。我们将这种不对称性归因于注意力转移（Attention Diversion）：词汇破坏打碎了子词分词，由此产生的碎片吸引了不成比例的注意力集中在Transformer的中层和末层。长度匹配的对照实验证实，是碎片化而非提示词长度导致了性能损失。随后的析因干预表明了为何这种损害难以逆转：碎片化同时破坏了分词内容与注意力分配，且两者紧密耦合。在内容仍然受损的情况下恢复干净的注意力反而会有害，仅恢复内容则远远不够，只有同时恢复两者才能弥补大部分性能差距。这种耦合解释了为什么推理阶段的策略（包括思维链提示、拼写检查、自我修复和更强的修复模型）无法持续恢复性能：因为它们每次只能处理其中一个通道。

Full-Text & External Resources

Full-Text & External Resources