Virginia Commonwealth University
Learning to Deceive and Persuade: Predictive Agents with Language and Personality Diversity in Social Deduction Games
Abstract
dc:description.abstractThis thesis presents a novel multi-agent architecture that combines Joint Em bedding Predictive Architectures (JEPA), Free Energy Principle (FEP)–inspired planning, and language-based social interaction. The system is evaluated in a custom Werewolf-style hidden-role game, where agents must coordinate, persuade, or deceive under partial observability. Each agent encodes its belief state into a latent representation, predicts future trajectories under candidate actions, and selects votes or eliminations through a planner trained to minimize prediction error. To enable communication, agents are equipped with trainable mouthpieces. A lightweight SpeakerBandit selects from discrete templates using REINFORCE on judge-derived rewards, while a logit-bias head steers a pretrained large language model by adjusting token probabilities across speech-act categories. Both mechanisms adapt over time, producing communication strategies aligned with role-specific goals. A rubric-driven LLM-as-Judge evaluates utterances for coherence, truthful viii ness, role alignment, and social safety, providing structured feedback that shapes both message generation and action selection. Agent diversity is introduced through personality traits sampled from config urable distributions. Personality knobs introduce behavioral variance by modulating speech-act preferences and style through fixed, sampled traits. Together, JEPA-based predictive modeling, FEP-style uncertainty minimization, judge-mediated language, and personality-driven variability yield agents that outperform reinforcement learning baselines in prediction accuracy, voting success, and communicative adaptability. The results demonstrate that predictive world models augmented with language and social influence provide a promising path toward more general and socially capable AI systems
Degree
thesis:*- Name thesis:degree_name
- Master of Science
- Level thesis:degree_level
- Thesis
- Discipline thesis:degree_discipline
- Computer Science
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Shah, Aaron M
- https://orcid.org/0009-0007-7605-424X
- Contributors dc:contributor
-
- Bridget McInnes
Subjects
dc:subject × 6Rights
dc:rights- Statement dc:rights
-
- © The Author
Identifiers
dc:identifier.*- Identifier
- https://scholarscompass.vcu.edu/etd/8234