Reproduce “Breaking the Assistant Mold: Modeling Behavioral Variation in LLM Based Procedural Character Generation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Maan Qraitem, Kate Saenko, Bryan A. Plummer

Boston University researchers developed PersonaWeaver, a framework that enhances the behavioral diversity of procedurally generated characters by large language models (LLMs), addressing inherent positive moral and helpful assistant biases. The approach disentangles character attributes into world-building and behavioral components, utilizing curated banks of moral stances and interactional styles. This method successfully generates character populations with a broader spectrum of moral dispositions, more varied conversational reactions, and emergent stylistic diversity across multiple LLMs.

Key insights

  • Large Language Models (LLMs) inherently possess alignment-induced biases (positive moral and helpful assistant) that lead to predictable and homogenized character outputs in PCG.
  • Explicitly disentangling and guiding LLM character generation with structured behavioral banks can successfully counteract these biases, revealing the models' latent capacity for diverse dispositions.
  • Introducing behavioral variation also induces emergent second-order stylistic diversity in character responses, such as varied use of filler words, punctuation, answer length, and sentiment distribution.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.