Breaking the Assistant Mold: Modeling Behavioral Variation in LLM Based Procedural Character Generation

TL;DR

The paper tackles the problem that LLM-based procedural character generation tends to produce biased, homogeneous agents due to alignment objectives. It introduces PersonaWeaver, a disentangled architecture that separately models world-building attributes and universal behavioral traits, enabling controlled variation via external banks. Behavioral banks based on Moral Foundations Theory for moral stances and an interactional taxonomy for conversational styles, plus a Sample and Mix mechanism, produce diverse character profiles. Across 10 settings and multiple frontier models, the approach yields broader distributions of moral positions, more varied reactions to questions, and measurable second-order stylistic diversity such as length, punctuation, and sentiment. The work demonstrates that decoupling world and behavior improves expressiveness in PCG and suggests a path toward richer, more credible social actors in virtual worlds, while acknowledging limitations and ethical considerations.

Abstract

Procedural content generation has enabled vast virtual worlds through levels, maps, and quests, but large-scale character generation remains underexplored. We identify two alignment-induced biases in existing methods: a positive moral bias, where characters uniformly adopt agreeable stances (e.g. always saying lying is bad), and a helpful assistant bias, where characters invariably answer questions directly (e.g. never refusing or deflecting). While such tendencies suit instruction-following systems, they suppress dramatic tension and yield predictable characters, stemming from maximum likelihood training and assistant fine-tuning. To address this, we introduce PersonaWeaver, a framework that disentangles world-building (roles, demographics) from behavioral-building (moral stances, interactional styles), yielding characters with more diverse reactions and moral stances, as well as second-order diversity in stylistic markers like length, tone, and punctuation. Code: https://github.com/mqraitem/Persona-Weaver

Figures (7)

• Figure 1: Illustration of biases in prior work versus our method Prior work populations jinworldweaver in (a)-Top collapse into uniform agreement with moral statements from the Social Chemistry dataset forbes-etal-2020-social, reflecting a moral bias, while in (a)-Bottom they invariably answer conversational questions from ConvAI2 dinan2019second, reflecting a reaction bias. In contrast, our generated populations (b)-Top and (b)-Bottom exhibit a broader range of stances and responses.
• Figure 2: Overview of PersonaWeaver, which disentangles world-building (a) from behavioral modeling (b) to give explicit control over behavioral variation (Moral Attitudes in this instance). A final Sample and Mix step (c) then combines these components into character profiles and ensures variation. Refer to Section \ref{['sec:personaweaver']} for further discussion.
• Figure 3: Comparison of Moral Positions between prior work WorldWeaverjinworldweaver, PersonaHubge2024scaling and our method PersonaHub. Refer to Sec \ref{['sec:bias_id']} and \ref{['sec:method_results']} for discussion.
• Figure 4: Comparison of Reactions to Questions between prior work WorldWeaverjinworldweaver, PersonaHubge2024scaling and our method PersonaHub. Refer to Sec \ref{['sec:bias_id']} and \ref{['sec:method_results']} for discussion.
• Figure 5: Comparison of Stylistic Patterns in the generated answers of prior work (WorldWeaverjinworldweaver and PersonaHubge2024scaling) and our work PersonaWeaver across four Stylistic categories (Filler words, Punctuations, Answer Length, and Sentiment). Refer to Section \ref{['sec:method_results']} for further discussion.