The Effect of Belief Boxes and Open-mindedness on Persuasion

The Effect of Belief Boxes and Open-mindedness on Persuasion Onur Bilgin1, Abdullah As Sami1, Sriram Sai Vujjini1 and John Licato1 1Advancing Machine and Human Reasoning (AMHR) Lab, University of South Florida Keywords: Belief-box, Open-mindedness, Multi-agent debate, Persuasiveness, Peer Pressure. Abstract: As multi-agent systems are increasingly utilized for reasoning and decision-making applications, there is a greater need for LLM-based agents to have something resembling propositional beliefs. One simple method for doing so is to include statements describing beliefs maintained in the prompt space (in what we’ll call their “belief boxes”). But when agents have such statements in belief boxes, how does it actually affect their behaviors and dispositions towards those beliefs? And does it significantly affect agents’ ability to be persuasive in multi-agent scenarios? Likewise, if the agents are given instructions to be open-minded, how does that affect their behaviors? We explore these and related questions in a series of experiments. Our findings confirm that instructing agents to be open-minded affects how amenable they are to belief change. We show that incorporating belief statements and their strengths influences an agent’s resistance to (and persuasiveness against) opposing viewpoints. Furthermore, it affects the likelihood of belief change, particularly when the agent is outnumbered in a debate by opposing viewpoints, i.e., peer pressure scenarios. The results demonstrate the feasibility and validity of the belief box technique in reasoning and decision-making tasks. 1 INTRODUCTION Argumentation is an important component of the reasoning and decision-making process. It al- lows individuals to communicate viewpoints, jus- tifications, and evidence. But there is a com- plex relationship between argumentation and be- lief. For example, individuals who hold beliefs strongly may be less receptive to arguments that go against those beliefs. Multi-agent large lan- guage model-based (LLM) systems have shown remarkable capabilities in various tasks such as problem-solving, decision-making, and reasoning (Qian et al., 2024; Li et al., 2023; Xu et al., 2024; Xiong et al., 2023; Xu et al., 2023). But understanding the beliefs of these agents (or the LLM-based analogue) is difficult, as the disposi- tions of agents are encoded in the distribution of weights in their neural architectures. It can there- fore be convenient (in terms of control, configu- ration, and explainability) to include an agent’s beliefs as explicit text that is provided in its in- put prompt. But how does the inclusion of beliefs in this way actually affect the individual and so- cial behaviors of these agents? For example, in environments where multiple agents with differ- ing beliefs interact, do belief boxes affect agents’ abilities to persuade each other, or to influence each other via peer pressure? We set out to explore answers to these ques- tions. In this paper, we define a construct that modifies an agent’s convictions about a natural- language proposition, leading the agent to adopt and defend it in a debate. We represent it with a belief box,1 a set of beliefs as propositions, each with a Likert scale strength value indicating the agent’s confidence in that belief. Hereafter, refer- ences to agents’ beliefs (e.g., aligned/misaligned or correct/incorrect) refer to the contents of the belief box, and we aim to evaluate whether the belief box aligns with the expected behavior of agents. Correct/incorrect refer to objective la- bels, while aligned/misaligned reflect beliefs cre- ated from human-annotated arguments in corre- sponding datasets. We explored beliefs in multi- agent systems, how the agents with beliefs in- teract with each other, how they evaluate and 1We borrow this term from philosophy of mind and cognitive science, where it often describes a common assumption of computationalism (Rescorla, 2019; Schiffer, 1981). arXiv:2512.06573v1 [cs.AI] 6 Dec 2025 change their beliefs, and how persuasive the group dynamics are. Contributions. Our research questions (RQs), hypotheses, and summaries of our results are listed in Table 1. We contribute to the existing literature by: • Providing the first (to our knowledge) de- tailed analysis of how equipping LLM-based agents with belief boxes affects their argumen- tation, belief change, and persuasion dynam- ics in groups. • Developing a multi-agent system framework for topic-driven debate simulation between agents with a belief box and a belief evaluation mechanism, drawing from the Aporia debate structure (Marji and Licato, 2021). • Providing evidence that: – assigning agents different levels of open- mindedness results in varying rates of belief change. – beliefs in agents’ belief boxes affect their ability to be persuasive about those beliefs. – the peer pressure effect exists; however, the extent of the effect varies with the group size. 2 RELATED WORK 2.1 Belief, Open-mindednes

This content is AI-processed based on open access ArXiv data.