Invited Talk 3 by Nazneen Rajani (CEO at Collinear AI, former Research Lead at Hugging Face): TraitBasis: An NPC Trait Engine for High-Fidelity RL Gym Simulations

Real-world AI agents need realistic NPCs (non playing characters)—diverse, expressive human personas that make RL environments feel alive and truly representative of how people interact. TraitBasis provides this missing layer, offering a lightweight, model-agnostic way to simulate controllable human traits that enrich training and evaluation. By bringing authentic behaviors into RL gyms, TraitBasis helps agents learn to adapt, generalize, and thrive under natural human variability. With τ-Trait expanding TauBench with airline, retail, telecom, and telehealth domains, TraitBasis opens the door to building AI agents that are not just robust, but genuinely ready for the real world.