Explicit Trait Inference for Multi-Agent Coordination

A team of researchers (Abdurahman, Ishii, Margatina, et al.) introduced Explicit Trait Inference (ETI), a lightweight method that enables LLM-based agents in multi-agent systems to infer and track partner characteristics along two psychological dimensions: warmth (e.g., trustworthiness) and competence (e.g., skill). By analyzing interaction histories, agents build structured trait profiles that guide decisions, reducing coordination failures like goal drift, error cascades, and misaligned behaviors.

In controlled economic games, ETI reduced payoff loss by 45-77% compared to baselines. In more realistic settings (MultiAgentBench), it improved performance by 3-29% depending on the scenario and model, relative to a chain-of-thought baseline. Additional analysis confirmed that gains are directly linked to trait inference accuracy: ETI profiles reliably predict agents' actions, and more informative profiles drive larger improvements. This provides the first systematic evidence that LLM agents can infer others' traits from interactions and use that awareness for coordination.