Simulating Hashtag Dynamics with Networked Groups of Generative Agents

ASONAM · Best Paper Runner Up · 2025

Jha, A., Priniski, J. H., Steinle, C., & Morstatter, F. Simulating hashtag dynamics with networked groups of generative agents. International Conference on Advances in Social Networks Analysis and Mining, 151–165.

Benchmarking AI Groups Against Human Communication

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).

The Hashtag Game with Networked LLM Agents

We prompted DeepSeek-R1, LLaMA-3, Qwen2, and Gemma agents to coordinate hashtags about the Fukushima nuclear disaster or Philippine presidential election. Agents interacted in Watts–Strogatz small-world networks, with prompts updated each round to include their neighbors’ previous hashtags.

For the human experiments used as a benchmark, see how network structure shapes consensus and narrative content in the Hashtag Game.

Comparing the Onset of Consensus in AI and Human Groups

We compared AI-agent communication dynamics with human dynamics to rank the models by their onset of consensus during networked interaction.

Qualitative ranking of consensus onset relative to human groups in Study 1 (Fukushima).
Rank	Model	Onset of consensus relative to human groups
1 (tie)	Gemma	Moved toward shared hashtags more slowly than human groups.
1 (tie)	Qwen2	Moved toward shared hashtags more slowly than human groups.
3	LLaMA-3	Showed little movement toward shared hashtags.
4	DeepSeek-R1	Responses became more varied rather than converging on shared hashtags.

Our simulations show that LLM-based agents can exhibit partial convergence in networked coordination tasks, but often fall short of replicating the onset of group coherence observed in human groups.

Shared Hashtags Can Flatten a Narrative’s Causal Structure

Number of human and LLM-generated hashtags most similar to each narrative entity in the Fukushima disaster. — **Figure 5.** Number of hashtags most similar to each narrative entity. While humans produced hashtags that aligned with both semantic topics (e.g., Fukushima Disaster) and cause-and-effect relationships, each of the LLMs produced almost exclusively topic-oriented hashtags.

Human participants generated hashtags that captured both semantic topics (e.g., Fukushima Disaster) and cause-and-effect relationships (e.g., Earthquake, Displacement). In contrast, the LLMs predominantly produced topic-oriented hashtags. These hashtags functioned as superficial topic labels for the event and did not encode the central causal relationships that human groups’ hashtags captured.

A core aim of my research is to investigate how to integrate the strengths of human reasoning with the linguistic coordination abilities of AI agents to help hybrid groups discover new forms of consensus. My ongoing human–AI group experiments extend this work by testing dynamically evolving agents that infer human belief states from language and adapt their messages during communication.

Political Narratives Present a Harder Coordination Test

In the second simulation, we examined political discourse coordination by simulating LLM-based agent behavior during the 2022 Philippine presidential election. None of the models demonstrated strong coordination, with only 10–50% of agents producing the same response.

Further inspection of the content of top hashtags revealed that models disproportionately focused on the first candidate listed in the prompt. To improve coordination and fairness in simulated political discourse, future prompting strategies should consider randomizing candidate order or explicitly instructing agents to sample across available options.