On the Persistent Effects of Lexicality in Large Language Models

关于大型语言模型中词汇性持续影响的研究

Abstract: Representations extracted from large language models (LLMs) play an important role in many downstream applications. However, the structure of these representations is often influenced by lexical overlap rather than semantic content. Our understanding of the relationship between this lexical influence and semantic content, and its implications for downstream tasks, remains limited.

摘要： 从大型语言模型（LLMs）中提取的表征在许多下游应用中发挥着重要作用。然而，这些表征的结构往往受到词汇重叠而非语义内容的影响。目前，我们对于这种词汇影响与语义内容之间关系及其对下游任务影响的理解仍然有限。

In this work, we investigate representations to quantify the effect of lexical overlap relative to semantic content. We consider several adversarial semantic stress tests and further connect our findings to the information theory perspective.

在这项工作中，我们通过研究表征来量化词汇重叠相对于语义内容的影响。我们考虑了多种对抗性语义压力测试，并将我们的研究结果与信息论视角进行了关联。

We find that lexical influence extends across the depth of models, consistently across architectures, training regimes, and objective functions, including the models trained for semantic similarity. Moreover, we observe a mid-depth region in which both lexical and semantic signals degrade simultaneously, indicating a transitional regime where representations are poor for both surface form and meaning.

我们发现，词汇影响贯穿于模型的整个深度，且在不同的架构、训练机制和目标函数（包括为语义相似度训练的模型）中表现一致。此外，我们观察到一个中间深度区域，在该区域中词汇信号和语义信号同时衰减，这表明存在一个过渡状态，使得表征在捕捉表面形式和深层含义方面表现均不佳。

We further demonstrate the effect of lexical influence on downstream uses of LLMs using summarization and model editing as a case study.

我们进一步以摘要生成和模型编辑为例，论证了词汇影响对大型语言模型下游应用的影响。

---

Paper Details:

• Authors: Hammad Rizwan, Muhammad Umair Haider, Nishant Subramani, Mona T. Diab, A.B. Siddique, Hassan Sajjad
• arXiv ID: 2606.02750
• Subject: Computation and Language (cs.CL)
• Submission Date: 1 Jun 2026

论文详情：

• 作者： Hammad Rizwan, Muhammad Umair Haider, Nishant Subramani, Mona T. Diab, A.B. Siddique, Hassan Sajjad
• arXiv ID: 2606.02750
• 学科： 计算与语言 (cs.CL)
• 提交日期： 2026年6月1日