Scholarlink

<p>Recent advancements in large language models (LLMs) have sparked growing interest in their ability to simulate not just linguistic competence, but agent like behavior. Moving beyond chatbots and task based assistants, researchers are now exploring whether LLMs can act as “single cognitive agents” entities endowed with memory, persona, planning capabilities, and situatedness in dynamic environments. This article presents the first structured literature review focused specifically on LLMs as single agent simulators. We examine over 70 peer reviewed publications and preprints (2018-2025), organizing them across five core dimensions: memory mechanisms, identity and persona modeling, planning and autonomy, interaction environments, and evaluation protocols. For each dimension, we analyze existing architectures, synthesize taxonomies, and identify recurring limitations. Our findings reveal a fragmented research landscape, where cognitive coherence, temporal memory, and personality persistence remain open challenges. We conclude by proposing a set of key research questions to guide future efforts, including the simulation of theory of mind, self consistent goal formation, and integrated memory persona frameworks. This review lays the groundwork for a unified research agenda aimed at developing cognitively coherent, memory augmented LLM based agents.</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

From prompt to persona: a literature review on LLMs as single cognitive agents

  • 期刊论文

  • 发表日期：

    2025年12月12日

查看全文
  • Anis Tissaoui

摘要

Recent advancements in large language models (LLMs) have sparked growing interest in their ability to simulate not just linguistic competence, but agent like behavior. Moving beyond chatbots and task based assistants, researchers are now exploring whether LLMs can act as “single cognitive agents” entities endowed with memory, persona, planning capabilities, and situatedness in dynamic environments. This article presents the first structured literature review focused specifically on LLMs as single agent simulators. We examine over 70 peer reviewed publications and preprints (2018-2025), organizing them across five core dimensions: memory mechanisms, identity and persona modeling, planning and autonomy, interaction environments, and evaluation protocols. For each dimension, we analyze existing architectures, synthesize taxonomies, and identify recurring limitations. Our findings reveal a fragmented research landscape, where cognitive coherence, temporal memory, and personality persistence remain open challenges. We conclude by proposing a set of key research questions to guide future efforts, including the simulation of theory of mind, self consistent goal formation, and integrated memory persona frameworks. This review lays the groundwork for a unified research agenda aimed at developing cognitively coherent, memory augmented LLM based agents.