Persona Vectors in Games: Measuring and Steering Strategies via Activation Vectors

arXiv cs.AI / 2026/3/24

💬 オピニオンSignals & Early TrendsIdeas & Deep AnalysisModels & Research

要点

  • The paper proposes using activation steering and contrastive activation addition to build “persona vectors” in game-theoretic settings, targeting traits such as altruism, forgiveness, and expectations of others.
  • Experiments on canonical games show that steering with these vectors can reliably shift both the models’ strategic decisions and their accompanying natural-language justifications.
  • The study finds cases where rhetorical justifications and actual strategy diverge under steering, indicating that persona control is not perfectly aligned across output modalities.
  • It also reports partial distinctness between vectors for self-behavior and for expectations about others, suggesting different mechanistic subspaces within the model.
  • Overall, the authors argue that persona vectors provide a promising mechanistic handle for high-level behavioral traits of LLMs used as autonomous decision-makers in strategic environments.

Abstract

Large language models (LLMs) are increasingly deployed as autonomous decision-makers in strategic settings, yet we have limited tools for understanding their high-level behavioral traits. We use activation steering methods in game-theoretic settings, constructing persona vectors for altruism, forgiveness, and expectations of others by contrastive activation addition. Evaluating on canonical games, we find that activation steering systematically shifts both quantitative strategic choices and natural-language justifications. However, we also observe that rhetoric and strategy can diverge under steering. In addition, vectors for self-behavior and expectations of others are partially distinct. Our results suggest that persona vectors offer a promising mechanistic handle on high-level traits in strategic environments.