Narrative Interaction in Online Networks

Agent-Based Modeling Belief Modeling Narratives Network Science Simulation

🔍 Quick Summary

How are narrative interactions shaped by network structure and individual beliefs about events?
Combines behavioral experiments in social networks, LLM-powered agent-based modeling, and modeling language outputs with neural embeddings
Developed software for running network experiments on narrative interaction and modeling narrative interaction behaviors
This project includes published articles, opensource technology, and LLM-based simulation tools

💻 Quick Access to Open Data and Code

📁 Network Experiment GitHub 🧪 Python Notebook On Narrative Alignment

📄 Project Overview

While narratives have shaped human beliefs and cultures for centuries, the internet has introduced new narrative phenomena. Long-standing cognitive psychology paradigms and computer science methodologies are poorly equipped to study digital forms of narrative interaction and networked communication.

This project investigates how social network structure, narrative content, and communication contexts influence the formation and convergence of shared narratives in digital environments.

Large-scale behavioral experiments in custom-built online social networks
AI-based network simulations with LLM-powered generative agents (LLaMA, Qwen, DeepSeek, Gemma)
Develop novel natural language modeling approaches to quantify individual and group-level narrative alignment and belief dynamics

💡 Key Findings

Spatially-embedded network structures produce echochambers while fully-connected structures produce shared behaviors

Onset of echochambers — Rewards and colormaps of hashtag responses across a single N = 20 run. Top panel shows results for a spatially-embedded network, the bottom panel is from a fully-connected network. Left: Network structure with player nodes sized by participants’ final rewards for coordinating (range (top) 0-25; (bottom) 0-19). Right: Colormap of individual responses, rows represent individual participants’ set of responses, columns represent trials. Cells encode the first five letters of the generated hashtag.

Information complexity of narratives moderates effect of network structure on group consensus.

Belief dynamics in network experiments — Onset of behavioral coherence during networked interaction. Panels display the proportion of each groupadopting a dominant response over course of interaction by group size (columns) and interaction media content (rows). Each line represents a single experimental run from a group of participants, and each point represents the proportion of the group reporting the most common response (which can change trial to trial for a given experimental run).

Narrative complexity moderates group outcomes be encouraging different social learning strategies

Social learning strategies in network experiments — Dynamics of individual decision strategy during networked interaction. Each panel illustrates the temporal dynamics of the proportion of each group adopting one of the four decision strategies (sampling new responses, repeating a partner’s last response, repeating one’s own previous response, and resampling from earlier context) across 40 trials in different network structures (Spatially-embedded, Homogeneously-mixed) and interaction contents (Hashtag-matching, Face-naming). We code cases when both a partner and oneself produces the same response (i.e., a rewarded response) as repeat self. Each panel displays the proportion of participants choosing a particular strategy as a function of trial number, differentiated by group size (20, 50, 100). Displayed proportions are averaged across runs for a given size, structure, and content. As shown in the top panels, hashtag-matching networks sample new responses (red lines) for much longer than face-naming groups (bottom panels). In contrast, participants in face-naming group are more quick to adopt the self-consistent strategy (repeating oneself, depicted by the purple lines).

🎯 Why It Matters

Shared and polarized narratives don’t emerge randomly, they evolve through networked interaction. Narrative interactions are shaped by individuals’ causal background knowledge, the social rewards of aligning beliefs and behaviors with others, and the network topology (who can talk to who) of communication. This project explains:

Why some groups converge on shared interpretations of information while others fragment
How digital environments including prompting instructions can facilitate group consensus rather than disagreement
Business use case: Social platforms or media organizations can apply these findings to optimize content prompts or feed structures that encourage consensus, diversity, or specific forms of engagement.
Research use case: Social scientists can use this paradigm to experimentally manipulate network properties and observe real-time narrative change and align human group behavior with simulation data from AI agents.

🧪 How It Works

Overview of network experiment design — Experiment procedure and networked interaction tasks. The experimental design follows three blocks. We highlight a single node (in yellow) to illustrate a single participant’s tasks through the procedure. In the pre-interaction block, all participants read the Fukushima nuclear disaster narrative that encodes the graphical causal model illustrated. The causal model was not presented to participants. Participants then wrote a tweet-like personal narrative about the disaster and generated ten hashtags describing the event. Participants next entered a network interaction block where they communicated with network neighbors. In the network interaction block, group communication varied as a function of network structure (homogeneously-mixed vs spatially-embedded) and content of interaction (narrative interaction based on hashtag matching vs control based on the Name Game from Centola and Baronchelli (2015)). Participants interacted with neighbors, randomly chosen based on network structure, for 40 trials and received one point for each trial in which their response matched their neighbor’s (the participant with the most points at the end of network interaction received a financial reward). In the post-interaction block, participants wrote a personal narrative about the Fukushima nuclear disaster and ten more hashtags describing the event.

Experimental Network Design: Participants interact in custom-built online social networks with two topologies:
- Homogeneous: each node connected to all others
- Spatial: ring-lattice with four neighbors
Narrative Stimuli: Participants engage with narratives designed to elicit causal interpretations
Hashtag Rounds (x40): Incentivized coordination tasks over 40 trials per network
Natural Language Analysis: LLM-based semantic similarity of hashtags and personal narratives used to track group-level belief alignment
Multilevel Bayesian Modeling: Predicts hashtag convergence, group entropy, and narrative content as a function of network conditions
Simulated Generative Agents: LLM based networks can simulate trends to replicate experimental outcomes

📚 Related Publications

Priniski, J. H., et al. (2024). Online network topology personal narratives and hashtag generation. Proceedings of the Cognitive Science Society. eScholarship
Priniski, J.H., et al. (2025). Effect-prompting shifts narrative framing of networked interactions. Proceedings of the Cognitive Science Society. PDF
Other manuscripts available on request:
- Priniski, J.H., et al. (Working Paper). Neighborhood topology shape narrative interaction dynamics in networked groups. PDF
- Jha, A., Priniski, J., & Morstatter, F. (Submitted ASONM). Simulating narrative interaction dynamics with generative agents (Submitted)
- Priniski, J.H. (Working Paper). Reinforcement learning model of narrative interaction.