SHARP - Unlocking Interactive Hallucination via Stance Transfer in Role-Playing LLMs

Aug 22, 2026

Upload by Aria

Role-playing capabilities of large language models (LLMs) enable rich multi-turn scenarios, while prior social-interaction research overlooks hallucination and often suffers from poor generalizability and implicit character-fidelity judgments. This work introduces a generalizable paradigm centered on stance transfer to explicitly define interactive hallucination. SHARP is built by extracting relations from commonsense knowledge graphs and leveraging LLM hallucination behavior to simulate multi-role interactions, with experiments validating effectiveness, stability, influencing factors, and exposing limits of common post-training mitigations.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages28

Reading Time1 h 11 min

AccessFree to download

Introduction

Interactive hallucination via stance transfer

SHARP benchmark construction

Experimental validation and analysis

arXiv :2411 .07965v 5 [ cs .CL] 3 Jun 2025 SHARP: Unlocking Interactive Hallucination via Stance Transfer in Role-Playing LLMs Chuyi Kong 1 , Ziyang Luo 1 , Hongzhan Lin 1 , Zhiyuan Fan3 , Yaxin Fan2 , Yuxi Sun 1 , Jing Ma 1 * , 1 Hong Kong Baptist University 2 Soochow University 3 The Hong Kong University of Science and Technology [m](majing@hkbu.edu.hk)[ajing@hkbu.edu.hk](majing@hkbu.edu.hk) Abstract The advanced role-playing capabilities of Large Language Models (LLMs) have enabled rich interactive scenarios, yet existing research in social interactions neglects hallucination while struggling with poor generalizability and implicit character fidelity judgments. To bridge this gap, motivated by human behaviour, we introduce a generalizable and explicit paradigm for uncovering interactive patterns of LLMs across diverse worldviews. Specifically, we first define interactive hallucination through stance transfer, then construct SHARP, a benchmark built by extracting relations from commonsense knowledge graphs and utilizing LLMs’ inherent hallucination properties to simulate multi-role interactions. Extensive experiments confirm our paradigm’s effectiveness and stability, examine the factors that influence these metrics, and challenge conventional hallucination mitigation solutions. More broadly, our work reveals a fundamental limitation in popular post-training methods for role-playing LLMs: the tendency to obscure knowledge beneath style, resulting in monotonous yet humanlike behaviors—interactive hallucination. 1 Introduction Large Language Models (LLMs) have evolved into versatile agents with impressive role-playing capabilities. Persona-based LLMs enhance reasoning and decision-making (Xu et al., 2024a) in specific domains (Kong et al., 2024a,b) but mainly focus on upstream indirect utility. In real-life downstream applications, LLMs role-played as characters inimmersive virtual worlds, such as Animations, Comics (Wu et al., 2025), Games (Wang et al., 2023 ; Fan et al., 2024 ; Wu et al., 2024), Novels (ACGN) (Tan et al., 2021 ; Spangher et al., 2024), and their corresponding drama adaptations, have drawn attention for their interactive features, bridging NLP and social psychology. *the corresponding author. Figure 1: Harry Potter’s wavering stance towards high affection-and low affection-level roles. More cases are shown in Appendix M.1 . According to social psychology (Cropanzano and Mitchell, 2005 ; Tedeschi, 2013), most humans adopt behaviors based on social connections. Even LLMs acting as judges exhibit preferences for their own series of models (Wataoka et al., 2024) . However, this unfair yet realistic behavior can be leveraged as a tool for immersive ACGN settings. For example, in role-playing games, players need to gain higher affection levels from non-player characters (NPCs) to progress, requiring game designers to establish assessment rules for such interactions. Motivated by this observation, we aim to uncover how well role-playing LLMs can capture human interaction behavior, and can LLMs be utilized asan automatic evaluator in role-playing. Given that most LLMs employ alignment techniques, such as SFT and RL (Shea and Yu, 2023), which align backbone models with instructions, they inevitably pay the alignment tax-hallucinations (Huang et al., 2023) . For instance, when a user asks counterfactual questions along with their own opinion, the model tends to adopt the user’s stance and agree with them in a sycophantic manner (Wei et al., 2023 ; Sharma et al., 2023), which | Benchmark | Focus | Format | Source | Judge | Metric | Automatic? | Generalizable? | | --- | --- | --- | --- | --- | --- | --- | --- | | HPD | Individual
Affection | Binary Label
(Rule) | Human | GPT-4,
Human | Scale
(-10-10) | ✗ | ✗ | | SocialBench | Group
Preference | MCQ
(Role Interaction) | GPT-4 + Human | Reference | Accuracy | ✓ | ✗ | | SHARP | Individual
Affection | Open-QA
(Role Intera

FAQs

What problem does the paper address in role-playing LLM research?

It targets the gap where social-interaction studies neglect hallucination and struggle with generalizability and implicit character-fidelity judgments in role-playing settings.

How does the SHARP benchmark generate multi-role interactions?

It extracts relations from commonsense knowledge graphs, converts part of them into counterfactuals, injects questioners’ opinions, and uses LLM hallucination behavior to simulate multi-role dynamics.

What is the paper’s central definition of interactive hallucination?

Interactive hallucination is defined as stance shifts based on backbone model tendencies or factual expectations during multi-role interactions.

Aug 22, 2026

Upload by Aria