LLMs Exhibit Malleable Social And Temporal Preferences

Thomas R. Cook, Sophia Kazinnik, Zach Modig, and Nathan M. Palmer (January 2026) analyze latent preferences of large language models using dictator-style allocation games and a sequential McCall job-search environment. They find models often favor equal splits—suggesting stronger inequality aversion than humans—and show prompt framing and control vectors can steer models toward payoff-maximizing behavior, though steering is less reliable in complex, dynamic tasks.