Personality anchoring makes LLM agents 10x more socially successful

A new paper from Vahid Sadiri Javadi and colleagues adapts the CHARISMA framework to create what they call "personality anchoring" for LLM agents. By grounding agents in well-known movie characters and public figures, the team simulated over 1,010 dyadic conversations across diverse scenarios. The key variable was Agreeableness composition—how agreeable or disagreeable each agent was relative to its partner. Results show a monotonic relationship: the more agreeable both agents were, the higher their shared goal achievement. Homogeneous-agreeable pairs succeeded 62% of the time, whereas homogeneous-disagreeable pairs managed only 6%—a tenfold gap.

Beyond the headline numbers, the researchers conducted behavioral mediation analysis to understand how agreeableness drives outcomes. They found that cooperative strategy selection partially explains the success, but agreeableness still predicts results even when controlling for the dominant conversational strategy—suggesting additional pathways beyond observable behavior. Robustness checks confirmed high consistency (ICC=0.89) across repeated simulations, and personality expression remained stable across diverse contexts. This work provides a scalable, validated method for studying how personality traits influence social interactions in AI systems, with implications for virtual assistants, social robotics, and team dynamics simulations.