Simulating hashtag dynamics with networked groups of generative agents

TL;DR

This paper investigates how networked narratives influence belief formation and coordination by simulating hashtag dynamics with networks of generative LLM agents. It introduces the Hashtag Matching game and benchmarks four diverse LLMs (DeepSeek-R1, LLaMA-3, Qwen2, Gemma2) against controlled human data (Fukushima) and real-world Twitter discourse (Philippines 2022 election) within a Watts-Strogatz network. The study uses entropy, narrative alignment via embedding, and unigram perplexity to measure group coherence and linguistic plausibility, uncovering that humans show stronger convergence than LLMs and that prompt design and social-context integration critically shape outcomes. The findings highlight domain dependence in LLM coordination, with Fukushima eliciting more narrative-conscious responses while the election domain yields greater lexical variability, underscoring the need for structured prompting and better alignment for realistic modeling of group narrative coordination and consensus formation.

Abstract

Networked environments shape how information embedded in narratives influences individual and group beliefs and behavior. This raises key questions about how group communication around narrative media impacts belief formation and how such mechanisms contribute to the emergence of consensus or polarization. Language data from generative agents offer insight into how naturalistic forms of narrative interactions (such as hashtag generation) evolve in response to social rewards within networked communication settings. To investigate this, we developed an agent-based modeling and simulation framework composed of networks of interacting Large Language Model (LLM) agents. We benchmarked the simulations of four state-of-the-art LLMs against human group behaviors observed in a prior network experiment (Study 1) and against naturally occurring hashtags from Twitter (Study 2). Quantitative metrics of network coherence (e.g., entropy of a group's responses) reveal that while LLMs can approximate human-like coherence in sanitized domains (Study 1's experimental data), effective integration of background knowledge and social context in more complex or politically sensitive narratives likely requires careful and structured prompting.

Figures (9)

• Figure 1: Procedure of the hashtag game using the Fukushima nuclear disaster narrative (Study 1) as the focal narrative and a fully-connected network group. The Hashtag Game follows a two block design. In the first block, all members of a group read a shared narrative about an event termed a focal narrative. Note participants do not see the causal model structure presented here, but rather read a text-based narrative describing that causal structure. In the second block, participants engage in networked interaction where they are incentivized to produce matching hashtags characterizing the narrative with network neighbors.
• Figure 2: Hashtags generated by the Llama network at the first and final trial.
• Figure 3: Change in entropy of each groups set of responses. Lower values indicate a large shift towards shared hashtag responses.
• Figure 4: Rank-abundance Curves (RACs) of each LLMs ten most frequently generated hashtags in response to the Fukushima disaster narrative.
• Figure 5: Number of hashtags most similar to each narrative entity. While humans produce hashtags that align with both semantic topics (e.g., Fukushima Disaster) and cause and effect relationships, each of the LLMs nearly exclusively produced hashtags about topics.