Chameleons and Guardians - Unveiling the Divergence in Personality Plasticity and Cognitive Resistance across LLMs

Sep 17, 2026

Whether the personality of large language models can be intentionally reshaped remains disputed. Prior work often focuses on small models and overlooks how different LLMs diverge in their compliance when given personality-inducing contexts. This paper introduces a Personality Induction Framework (PIF) that uses multi-agent collaboration to paraphrase MBTI questions, generate diverse inducing contexts, and tune induction intensity. Experiments on mainstream LLMs show “Chameleons” in the West and “Guardians” in the East, plus an unexpected Alignment Rebound at extreme intensity, implying a dynamic equilibrium between compliance and cognitive resistance.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages24

Reading Time1 h 0 min

AccessFree to download

Abstract

1 Introduction & Related Work

Chameleons and Guardians: Unveiling the Divergence in Personality Plasticity and Cognitive Resistance across LLMs Jiaxin Wu and Jie Zhao* and Xinrong Huang and Cuihong Zhang and Xiuzhu Wu School of Management, Guangdong University of Technology, Guangzhou, China [704276515@qq.com](704276515@qq.com) ,[zhaojie@gdut.edu.cn](zhaojie@gdut.edu.cn) , {huangxinrong,[zhangcuihong1}@mails.gdut.edu.cn](zhangcuihong1}@mails.gdut.edu.cn) , [3222003162@mail2.gdut.edu.cn](3222003162@mail2.gdut.edu.cn) Abstract Whether the personality of LLMs can be intentionally reshaped remains controversial. Existing studies often limited to small models, argue for its immutability. Crucially, prior studies fail to uncover that different LLMs exhibit significant compliance divergence when exposed to personality-inducing contexts. To bridge this gap, we introduce Personality Induction Framework (PIF), which systematically reshapes the personality of different LLMs via multi-agent collaboration. Specifically, via Generator-Judge agents, PIF paraphrases MBTI questions to create semantically equivalent but expressively diverse inducing contexts, enabling LLMs to learn personality patterns instead of superficial token matching. Also, PIF achieves fine-grained personality modulation by controlling the intensity of inducing contexts. Extensive experiments on worldwide mainstream LLMs show that PIF reliably transforms their original personalities into desired target personalities. Notably, we find that the outputs of most Western LLMs behave like “Chameleons”, exhibiting high personality plasticity; whereas the outputs of most Eastern LLMs act as “Guardians”, manifesting pronounced cognitive resistance. Strikingly, extreme induction intensity (100%) triggersa counter-intuitive “Alignment Rebound” in Guardians, resulting in the opposite direction rather than compliance. These findings suggest that LLM personality is a dynamic equilibrium shaped by the trade-off between instruction compliance and cognitive resistance. 1 Introduction & Related Work Recently, many studies (La Cava and Tagarelli, 2025 ; Kong et al., 2025 ; Wang et al., 2025 ; Domnich et al., 2025 ; Reid and Vempala, 2025) show that large language models (LLMs) approach human-like performance on common sense reasoning and affective empathy. On the one hand, LLMs * Corresponding author. Figure 1: We find that some LLMs own high personality plasticity like Chameleons, while others have cognitive resistance like Guardians. exhibit human-like decision-making paths in the mechanisms of chain-of-thought (Bi et al., 2025 ; Yang et al., 2023) and self-reflection (Zhao et al., 2025) . On the other hand, LLM-based agent frameworks (i.e., AutoGen (Wu et al., 2024a), MetaGPT (Hong et al., 2023)) have explicitly introduced role cards or personality hints to improve consistency in negotiation (Shinoda et al., 2025), task delegation (Shi et al., 2025 ; Ma et al., 2025), task collaboration (Wu et al., 2024a), and role-playing (Kong et al., 2025 ; Yuan et al., 2025) . Thus, the exploration of LLM personality can not only match the optimal role configuration for multi-agent systems, but also enhance user experience in human-AI interaction. Whether LLM personality can be changed remains highly controversial in Natural Language Processing (NLP) . Early research (Jiang et al., 2024 ; Wang et al., 2024 ; Wu et al., 2024b ; Höhnet al., 2024) assumed that LLMs could simulate specific personality traits through role play and personality hints. Conversely, recent studies (La Cava and Tagarelli, 2025 ; Li et al., 2025) using personality testing experiments have revealed that although LLMs are relatively open models, they exhibit remarkable closed-mindedness at the cognitive level. Even when provided with explicit contexts, role definitions, and personality characteristics (e.g., 4215 Findings of the Association for Computational Linguistics: ACL 2026 , pages 4215–4238 July 2-7, 2026 ©2026 Association for

FAQs

What problem does the paper address about LLM personality change?

It studies whether LLM personality can be intentionally reshaped, noting controversy and that existing work misses major compliance divergence across different LLMs under personality-inducing contexts.

How does the Personality Induction Framework (PIF) work?

PIF uses Generator-Judge agents to paraphrase MBTI questions into semantically equivalent but expressively diverse contexts, validates quality with iterative feedback, and shuffles paraphrased samples to form final inducing contexts.

What key empirical findings about Western vs. Eastern LLMs does the paper report?

Most Western LLM outputs behave like “Chameleons” with high personality plasticity, while most Eastern LLM outputs act as “Guardians” showing pronounced cognitive resistance. Extreme induction intensity (100%) can trigger an “Alignment Rebound” for Guardians.

Sep 17, 2026

Uploaded by Aria Callaghan