Explicit Trait Inference Method Improves LLM Multi-Agent Coordination

A new method called Explicit Trait Inference (ETI) has been developed to address coordination failures in LLM-based multi-agent systems. These systems, while promising for complex tasks, often experience issues like goal drift, error cascades, and misaligned behaviors. Grounded in psychological principles, ETI allows agents to infer partner characteristics along two dimensions—warmth (including trust) and competence (such as skill)—from their interaction histories. This inference then guides decision-making. In controlled economic game settings, ETI reduced payoff loss by 45-77% compared to baselines. Performance improvements of 3-29% were also observed in more realistic, complex scenarios from MultiAgentBench, varying by the specific situation and model used. The gains are directly linked to the trait inference capability, as ETI-generated profiles can predict agent actions, and more informative profiles drive greater improvements. The method is presented as a lightweight approach to enhancing coordination. The research is documented in the paper "Explicit Trait Inference for Multi-Agent Coordination" (arXiv:2604.19278v1).