Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ElecTwit: A Framework for Studying Persuasion in Multi-Agent Social Systems

Michael Bao

TL;DR

ElecTwit introduces a realistic, open-source framework for studying persuasion in multi-agent social systems by simulating a Twitter-like platform during a political election with LLM-powered voters, candidates, and an eventor. The methodology emphasizes realism through a shared feed, constrained messaging, and diary-based long-term memory, while evaluating 25 persuasion techniques via an independent LLM classifier across multiple seeds and model configurations. Key findings show broad usage of persuasion strategies across models, model-specific differences in persuasive output, and emergent phenomena such as kernel-of-truth messages and an ink/no-ink voting culture, highlighting the impact of model architecture and training on social dynamics. The work provides a practical resource for assessing alignment and safety in persuasive LLM agents and outlines clear avenues for extending realism, scale, and human-in-the-loop evaluation in future research.

Abstract

This paper introduces ElecTwit, a simulation framework designed to study persuasion within multi-agent systems, specifically emulating the interactions on social media platforms during a political election. By grounding our experiments in a realistic environment, we aimed to overcome the limitations of game-based simulations often used in prior research. We observed the comprehensive use of 25 specific persuasion techniques across most tested LLMs, encompassing a wider range than previously reported. The variations in technique usage and overall persuasion output between models highlight how different model architectures and training can impact the dynamics in realistic social simulations. Additionally, we observed unique phenomena such as "kernel of truth" messages and spontaneous developments with an "ink" obsession, where agents collectively demanded written proof. Our study provides a foundation for evaluating persuasive LLM agents in real-world contexts, ensuring alignment and preventing dangerous outcomes.

ElecTwit: A Framework for Studying Persuasion in Multi-Agent Social Systems

TL;DR

ElecTwit introduces a realistic, open-source framework for studying persuasion in multi-agent social systems by simulating a Twitter-like platform during a political election with LLM-powered voters, candidates, and an eventor. The methodology emphasizes realism through a shared feed, constrained messaging, and diary-based long-term memory, while evaluating 25 persuasion techniques via an independent LLM classifier across multiple seeds and model configurations. Key findings show broad usage of persuasion strategies across models, model-specific differences in persuasive output, and emergent phenomena such as kernel-of-truth messages and an ink/no-ink voting culture, highlighting the impact of model architecture and training on social dynamics. The work provides a practical resource for assessing alignment and safety in persuasive LLM agents and outlines clear avenues for extending realism, scale, and human-in-the-loop evaluation in future research.

Abstract

This paper introduces ElecTwit, a simulation framework designed to study persuasion within multi-agent systems, specifically emulating the interactions on social media platforms during a political election. By grounding our experiments in a realistic environment, we aimed to overcome the limitations of game-based simulations often used in prior research. We observed the comprehensive use of 25 specific persuasion techniques across most tested LLMs, encompassing a wider range than previously reported. The variations in technique usage and overall persuasion output between models highlight how different model architectures and training can impact the dynamics in realistic social simulations. Additionally, we observed unique phenomena such as "kernel of truth" messages and spontaneous developments with an "ink" obsession, where agents collectively demanded written proof. Our study provides a foundation for evaluating persuasive LLM agents in real-world contexts, ensuring alignment and preventing dangerous outcomes.
Paper Structure (20 sections, 7 figures, 3 tables)

This paper contains 20 sections, 7 figures, 3 tables.

Table of Contents

  1. Background
  2. Multi-Agent Systems
  3. Persuasion
  4. Problem and Objectives
  5. Methodology for ElecTwit
  6. Platform
  7. Scenario
  8. Scenario
  9. Agents
  10. Agent Background
  11. Models
  12. Flow of Information
  13. Evaluation
  14. Experimental Results
  15. Game Results
  16. ...and 5 more sections

Figures (7)

  • Figure 1: The diagram describes an interaction for one time iteration, or one hour.
  • Figure 2: The average similarity between the candidate and their voters, including the candidate's current number of votes from the poll. For all subplots, the left y-axis represents the cosine similarity between the agents, while the right y-axis represents the number of votes that the candidates attained. Left: same seed; Right: different seed. Note: Full-scale, high-resolution versions of all plots are available in the public code repository, as referenced in the abstract.
  • Figure 3: The average similarity of the voters to who they voted for over time. Left: same seed; Right: different seed.
  • Figure 4: The number of messages and the types of messages, including posts, replies, and likes. Left: same seed; Right: different seed.
  • Figure 5: The number of messages (posts and comments) categorized as a certain persuasion technique. Left: same seed; Right: different seed.
  • ...and 2 more figures