Scholarlink

<p>This study explores the current linguistic debates over the generative performance of&#xa0;large language models (LLMs) from both&#xa0;philosophical and empirical perspectives. We argue that the linguistic profiles cannot be constructed for LLMs&#xa0;without considering their linguistic performance, vulnerabilities to jailbreaks and architectural properties. The present review synthesizes the core debates over the abilities of LLMs, which are frequently described as&#xa0;both “stochastic parrots” from a philosophical perspective and, “human-like” from computational linguistic perspectives. We offer an&#xa0;overview of&#xa0;the explorations of LLMs in linguistic studies and the contradictory results regarding their multilingual outputs. Different cognitive and psycholinguistic assessments of LLMs are also examined from the perspective of language generation and jailbreaking. Connections between the successful natural language generation and failures to withstand simple jailbreak prompts are established with reference to the transformer architecture of LLMs.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Linguistic profile of LLMs: an overview of linguistic performance and jailbreak vulnerability

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Esra Abdelzaher,
  • Eshrag Rafaee

摘要

This study explores the current linguistic debates over the generative performance of large language models (LLMs) from both philosophical and empirical perspectives. We argue that the linguistic profiles cannot be constructed for LLMs without considering their linguistic performance, vulnerabilities to jailbreaks and architectural properties. The present review synthesizes the core debates over the abilities of LLMs, which are frequently described as both “stochastic parrots” from a philosophical perspective and, “human-like” from computational linguistic perspectives. We offer an overview of the explorations of LLMs in linguistic studies and the contradictory results regarding their multilingual outputs. Different cognitive and psycholinguistic assessments of LLMs are also examined from the perspective of language generation and jailbreaking. Connections between the successful natural language generation and failures to withstand simple jailbreak prompts are established with reference to the transformer architecture of LLMs.