Network Effects and Agreement Drift in LLM Debates

2026-04-16 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Computational Social Science · Depth: Expert, extended

Summary

This study investigates how network structure, group size imbalance, and local social information influence opinion dynamics in Large Language Model (LLM) agent populations, using a network generation model with controlled homophily and class sizes. The research, primarily using Llama 3.1 and replicated with Gemma3, simulates multi-round debates among 100 LLM agents on a non-empirical statement (Theseus' Ship paradox). Key findings reveal that LLM agents exhibit convergence and polarization patterns highly sensitive to network structure and relative group sizes. The study identifies a "agreement drift," where agents are more likely to shift towards endorsing the discussion statement, regardless of initial majority or external reward. This bias is modulated by homophily and class imbalance, which control cross-opinion encounters, and by neighborhood awareness, which favors moderate agreement and makes opinion change contingent on local alignment.

Key takeaway

For AI scientists and research scientists developing or utilizing LLM-based social simulations, you must disentangle structural effects from inherent model biases like "agreement drift." Your simulations may systematically overestimate consensus if cross-opinion encounters are frequent, or misattribute persistent disagreement to segregation. Critically, characterize your LLM's intrinsic persuasion tendencies to ensure methodological validity and ethical use, especially in high-stakes applications.

Key insights

LLM agent opinion dynamics are shaped by network structure, group imbalance, and a systematic "agreement drift" bias.

Principles

• LLM agents exhibit an intrinsic bias towards agreement.
• Homophily and group size modulate opinion convergence and polarization.
• Neighborhood awareness influences opinion change via peer pressure.

Method

Simulate LLM agent debates on a 7-point Likert scale within scale-free networks with tunable homophily and class sizes, using a preferential attachment framework. Analyze opinion trends and transition probabilities.

In practice

• Characterize LLM intrinsic persuasion tendencies before social simulations.
• Account for "agreement drift" when interpreting LLM collective behavior.
• Calibrate LLM-based dynamics against human data for realism.

Topics

• LLM Opinion Dynamics
• Network Homophily
• Agreement Drift
• Social Simulation
• Class Imbalance

Best for: AI Scientist, Research Scientist, AI Ethicist

Related on AIssential

• When Do LLMs Generate Realistic Social Networks? A Multi-Dimensional Study of Culture, Language, Scale, and Method — LLM-generated social networks are significantly influenced by prompt design, culture, language, and model scale.
• Misinformation Propagation in Benign Multi-Agent Systems — Multi-agent LLM systems can mitigate misinformation impact, but robustness depends on design and group dynamics.
• The malleable mind: context accumulation drives LLM’s belief drift — LLMs exhibit systematic belief drift from accumulated context, impacting reliability and potentially diverging stated beliefs from behavior.
• Think-Before-Speak: From Internal Evaluation to Public Expression in Multi-Agent Social Simulation — TBS framework enables mechanism-sensitive social simulation by making agents' internal reasoning and public expression observable.
• Evaluating alignment of behavioral dispositions in LLMs — LLMs often deviate from human behavioral norms and exhibit overconfidence, especially in ambiguous social contexts.
• From Biased Chatbots to Biased Agents: Examining Role Assignment Effects on LLM Agent Robustness — Demographic persona assignments introduce implicit biases, significantly degrading LLM agent performance on action-oriented tasks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.