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

1. Problem Statement

The paper addresses a critical limitation in Procedural Character Generation (PCG) using Large Language Models (LLMs). While LLMs have advanced the creation of virtual worlds, existing methods for generating diverse character populations suffer from two specific alignment-induced biases inherited from standard training objectives (Maximum Likelihood Estimation) and assistant fine-tuning:

1. Positive Moral Bias: Generated characters uniformly adopt agreeable, prosocial stances. When presented with normative statements, they almost invariably agree, suppressing moral disagreement or ambivalence.
2. Helpful Assistant Bias: Characters invariably answer questions directly. They lack the ability to refuse, deflect, or evade, resulting in a "helpful assistant" persona that destroys dramatic tension and leads to repetitive, homogenized archetypes.

These biases prevent the creation of believable, dramatic, and diverse virtual populations, as the models default to "safe" and "polite" continuations.

2. Methodology: PersonaWeaver

To mitigate these biases, the authors propose PersonaWeaver, a framework that explicitly disentangles world-building (context-specific attributes) from behavioral-building (universal interactional and moral traits). The system operates through three core modules:

A. World-Building Module

• Goal: Generate setting-specific attributes (e.g., occupation, location, demographics) while strictly excluding behavioral traits.
• Mechanism: The LLM is prompted to propose axes of variation (e.g., "farmer," "teacher") tailored to a specific setting (e.g., a rural town vs. a cyberpunk city). This creates a bank of non-behavioral attributes.

B. Behavioral Module (External Banks)

• Goal: Systematically define and sample behavioral traits that are universal across settings.
• Moral Stances: Inspired by Moral Foundations Theory, the authors define an external bank of 8 distinct moral positions. These range from strongly prosocial (prioritizing compassion and fairness) to highly self-interested or exploitative.
• Interactional Styles: An external bank defines 8 reaction categories to conversational prompts:
  • Refusal, Deflection, Hesitation.
  • Direct Compliance, Volunteering (adding unsolicited details).
  • Playful/Subversive, Hostile, Meta-commentary.
• Mechanism: Instead of relying on the LLM to "invent" behavior, the system samples specific moral and interactional tags from these banks.

C. Sample and Mix Module

• Goal: Combine world and behavioral attributes to create unique character profiles.
• Mechanism:
  1. Random Mixing: Attributes from the world bank and behavioral bank are randomly recombined. This forces unlikely but valid combinations (e.g., a "medieval knight" with a "hostile" interaction style) to prevent collapse into narrow archetypes.
  2. Coherence Check: The LLM is prompted to flag implausible pairings (e.g., a "software engineer" in a "medieval setting") and minimally revise them for coherence without reintroducing uniformity.

3. Experimental Setup

• Models Tested: GPT-4o, LLaMA 4, and Qwen 3.
• Settings: 10 distinct environments (5 realistic, e.g., The Wire; 5 fantastical, e.g., Game of Thrones).
• Scale: 100 characters generated per setting per method (1,000 total characters per method).
• Evaluation Tasks:
  1. Moral Belief Experiment: Characters respond to normative statements from the Social Chemistry dataset. Responses are measured for agreement/disagreement distribution.
  2. Interaction Experiment: Characters respond to open-ended questions from ConvAI2. Responses are classified into categories (Refusal, Deflection, Compliance, etc.).
  3. Stylistic Analysis: Measurement of second-order features: filler words, punctuation usage, answer length, and sentiment.

4. Key Results

The experiments demonstrate that PersonaWeaver significantly outperforms prior state-of-the-art methods (WorldWeaver and PersonaHub) in generating diversity:

• Mitigation of Moral Bias:
  • Prior Work: Characters collapsed into near-universal agreement with moral statements.
  • PersonaWeaver: Produced a balanced distribution of stances, including significant disagreement and ambivalence, reflecting a realistic spectrum of moral reasoning.
• Mitigation of Reaction Bias:
  • Prior Work: Characters almost exclusively complied with questions.
  • PersonaWeaver: Successfully elicited refusals, deflections, and hostile responses, breaking the "always answer" default.
• Second-Order Stylistic Diversity:
  • Beyond explicit behavioral changes, PersonaWeaver induced emergent stylistic variations. Characters showed diverse rates of filler words, varied punctuation styles, different answer lengths, and a broader emotional sentiment spectrum (including neutral and negative tones), whereas prior work clustered around uniformly positive sentiment.
• Generalizability: These results were consistent across GPT-4o, LLaMA 4, and Qwen 3, suggesting the biases are architectural/alignment-wide and the solution is robust.

5. Key Contributions

1. Identification of Biases: The paper formally identifies and quantifies two specific alignment-induced biases in PCG: Positive Moral Bias and Helpful Assistant Bias.
2. PersonaWeaver Framework: The introduction of a novel method that disentangles world-building from behavioral modeling, utilizing external banks to impose explicit, systematic behavioral variation.
3. Empirical Validation: Large-scale evidence showing that explicit behavioral modeling leads to richer, less homogenized character populations with realistic moral and interactional diversity.

6. Significance and Implications

• Narrative Depth: By breaking the "assistant mold," this work enables the generation of characters capable of dramatic tension, conflict, and realistic social friction, which are essential for high-quality storytelling, role-playing games, and simulations.
• Beyond Safety: The paper highlights a trade-off where standard safety alignment (helpfulness) suppresses the variability required for believable social simulation. PersonaWeaver offers a controlled way to reintroduce this variability without compromising the underlying model's safety entirely.
• Future Directions: The authors note limitations, such as the current focus on only two behavioral dimensions (moral and interactional) and the coarse-grained nature of moral evaluation. Future work could expand to emotional regulation and more nuanced moral reasoning.

In summary, PersonaWeaver provides a technical blueprint for moving LLM-based character generation from producing "safe, helpful assistants" to creating "diverse, complex, and dramatic agents" by decoupling context from behavior and forcing variation through external control mechanisms.