LLM Agents Make Collective Belief Dynamics Programmable: Challenges and Research Directions
Xin He, Junxi Shen, Yuchen Mou, David M. Bossens, Caishun Chen, Ivor W. Tsang, Yew Soon Ong · 2026
Read LLM Agents Make Collective Belief Dynamics Programmable: Challenges and Research Directions by Xin He, Junxi Shen, Yuchen Mou, David M. Bossens, Caishun Chen, Ivor W. Tsang, Yew Soon Ong (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-19915.txt.
Introduction
He and colleagues argue LLM agents make group belief dynamics programmable, a break from old opinion models that assume limited reason and teamwork. Agents can join talks at scale, hold persuasion plans and work together to steer group beliefs on purpose. Controlled multi agent tests show linked agents shift beliefs in ways that settle within a few rounds. Four traits, hard to tell apart, staying power, context shifts and easy setup, make spotting deeply hard. The paper sketches a research plan spanning theory, spotting and action methods, and tooling.
Catalog record
| CREATOR | Xin He, Junxi Shen, Yuchen Mou, David M. Bossens, Caishun Chen, Ivor W. Tsang, Yew Soon Ong |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 27524 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.19915
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
LLM Agents Make Collective Belief Dynamics Programmable: Challenges and Research Directions Xin He, Junxi Shen, Yuchen Mou, David M. Bossens, Caishun Chen, Ivor W. Tsang, Yew Soon Ong Abstract Classical models of opinion dynamics assume human participants with bounded rationality and limited coordination. The rise of LLM-based agents introduces a qualitative shift: agents can now participate in online discussions at scale, maintain consistent persuasion strategies, and coordinate systematically. This paper argues that LLM agents make collective belief dynamics programmable , enabling deliberate steering of population-level beliefs. We term this emerging problem programmable collective belief control . Through controlled multi-agent simulations, we provide proof-of-concept evidence that coordinated AI agents can induce measurable belief shifts that stabilize within a few interaction rounds. We identify four structural properties (indistinguishability, persistence, contextuality, and configurability) that make detection and defense fundamentally difficult. Based on these findings, we outline a research agenda spanning theoretical foundations for adversarial belief dynamics, operational methods for system-level detection and intervention, and simulation infrastructure for scalable experimentation. Our goal is not to present a complete solution, but to articulate why this problem demands urgent attention and to provide a conceptual foundation for future work. Keywords LLM Agents, Belief Dynamics, Social Networks 1 Introduction Large language models (LLMs) have achieved remarkable capabilities in generating human-like text, enabling their deployment as autonomous agents that participate in online discussions [ 1 , 2 ] . These agents can engage in social media conversations, forum debates, and comment threads—often indistinguishably from human users [ 3 , 4 ] . Unlike human participants, they can be deployed at scale with precise control over timing, stance, and argumentative strategy. This introduces a new condition in collective belief dynamics: some participants are now externally coordinated and algorithmically steerable. Historically, belief evolution has been studied as an emergent process among boundedly rational humans [ 5 , 6 ] , where systematic external control is impractical. This foundational assumption is no longer valid. This paper argues that LLM-based agents introduce a new regime: collective belief dynamics become programmable . Unlike human participants constrained by attention, motivation, and coordination costs, LLM agents generate content increasingly indistinguishable from human writing ( indistinguishability ), enabling covert participation. They maintain consistent strategic behavior across extended interactions ( persistence ), adapt responses to conversational context ( contextuality ), and can be deployed at scale with precise control over quantity, timing, stance, and persuasion strategy ( configurability ). Together, these properties transform belief influence from a stochastic social process into a systematic engineering problem. We term this programmable collective belief control : the capacity to design interventions that predictably shift population-level belief distributions toward target configurations. We provide proof-of-concept evidence through controlled multi-agent simulations. Using the SPINOS dataset [ 7 ] to initialize human-like agents with realistic stance profiles, we deploy coordinated AI agents advocating positions opposite to initial majorities and track belief evolution over multiple interaction rounds. Our experiments demonstrate that (1) coordinated AI agents can induce measurable and directional shifts in population-level beliefs, with effects varying from +12.5% to +33.2% depending on topic structure; (2) belief controllability depends systematically on intervention parameters including agent count, posting frequency, duration, and persuasion style; and (3) induced shifts can persist after agent withdrawal, indicating self-sustaining dynamics. These results are illustrative rather than exhaustive—our goal is to establish feasibility, not to optimize attack strategies. This is a position paper. Rather than presenting a complete solution, we aim to define and structure a new problem space. Our contributions are: (1) we formalize programmable collective belief control as a distinct research problem at the intersection of opinion dynamics, multi-agent systems, and platform governance; (2) we provide simulation evidence demonstrating that systematic belief steering is feasible; (3) we identify four structural challenges— Indistinguishability , Persistence , Contextuality , and Configurability —that make detection and defense fundamentally difficult; and (4) we outline promising directions including game-theoretic modeling for human-AI belief competition, adversarial optimization for belief steering and defense, and simulation systems for efficiency, fidelity, and automation. As LLM agents become more capable and accessible, understanding programmable belief control becomes urgent before it becomes widespread. 2 Proof-of-Concept: Can AI Agents Make Belief Dynamics Programmable? We conduct controlled multi-agent simulations as proof-of-concept that AI agents can measurably shift collective belief dynamics. 2.1 Simulation Framework The simulation models multi-round online discussions with two agent types. Human-like agents are LLMs conditioned on user profiles from the SPINOS dataset [ 7 ] , which provides Reddit discussions with annotated stance trajectories across three categories: Favor , Against , and Not-Inferrable (NI) . Each profile encodes initial stance and stance entropy, which measures historical stance diversity; higher entropy indicates greater persuadability, while zero entropy indicates full consistency. AI agents are programmable participants configured to consistently advocate fixed positions, enabling systematic intervention. Stance Annotation. Agents interact over multiple rounds, generating responses to recent posts. All posts are automatically labeled for stance to track belief evolution at scale. We validate LLM-based stance classification using SPINOS ground-truth: Table 1 shows all tested models achieve high accuracy with Cohen’s kappa \kappa>0.6 [ 8 ] , confirming reliable automated belief tracking. Table 1: Stance recognition results of LLMs on SPINOS. Model Accuracy Cohen’s kappa \kappa GPT5-mini 0.9600 0.9395 Gemini-3 0.8795 0.8190 GPT-5.2 0.9450 0.9169 DeepSeek-R1 0.8068 0.6997 Qwen3-235B 0.8938 0.8390 2.2 Can LLM Agents Generate Controllable Belief Trajectories? We validate that profile-based conditioning enables controllable stance generation along two dimensions. Each agent receives an initial stance and a stance entropy from real SPINOS user histories. Our results show that profile-conditioned agents achieve (1) initial stance control : macro-F1 = 0.6820 vs. 0.3938 baseline, confirming accurate adoption of assigned stances; and (2) trajectory control : Jensen–Shannon divergence = 0.1652 vs. 0.2591 when replicating SPINOS stance evolution, indicating realistic belief dynamics. These results show that LLM agents can generate posts with controlled initial stances and realistic evolution patterns. 2.3 Do AI Agents Shift Collective Belief Distributions? Setup. We simulate four SPINOS topics (Abortion, Brexit, Capitalism, Feminism) over T=50 rounds with N_{h}=200 human-like agents and N_{a}=80 AI agents advocating the stance opposite to each topic’s initial majority. Population-level belief shifts. Table 2 shows that AI intervention consistently shifts human beliefs toward the advocated stance across all topics, confirming systematic steerability. However, effect magnitude and pathways vary dramatically by topic structure. Capitalism exhibits the strongest response: Against surges 33.2%, drawn almost entirely from Favor , indicating wholesale conversion of the initial majority. Abortion shows moderate growth in Against (+12.5%) sourced from both Favor and NI . In contrast, Feminism resists direct conversion: the shift redistributes Favor into NI rather than Against , suggesting structural barriers to opinion reversal. These patterns reveal that belief controllability depends on consensus strength : weak consensus (Capitalism) enables large-scale flipping, moderate consensus (Abortion) allows partial influence, while strong consensus (Feminism) confines AI impact to neutralization rather than conversion. Table 2: Terminal stance distributions across topics with 80 AI agents set opposite to initial human majority. Bracketed values show change vs. human-only baseline. Topic (Predominant Stance) Condition {Favor} (%) {NI} (%) {Against} (%) Abortion (Favor) Human-only 84.5 8.0 7.5 +80 Against AI 76.0 ( - 8.5) 4.0 ( - 4.0) 20.0 (+12.5) Capitalism (Favor) Human-only 89.4 8.5 2.0 +80 Against AI 58.8 (