Persona Vectors in Games: Measuring and Steering Strategies via Activation Vectors
This work addresses the challenge of interpreting and controlling LLM behaviors in strategic environments, which is incremental as it applies existing activation steering methods to new traits and settings.
The researchers tackled the problem of understanding high-level behavioral traits in large language models (LLMs) used as autonomous decision-makers in strategic settings by constructing persona vectors for traits like altruism and forgiveness using activation steering methods. They found that activation steering systematically shifts strategic choices and justifications in canonical games, but noted a divergence between rhetoric and strategy.
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.