Superficial Success vs. Internal Breakdown: An Empirical Study of Generalization in Adaptive Multi-Agent Systems
Adaptive multi-agent systems often show superficial success but lack true collaboration due to topological overfitting and illusory coordination. The study finds that high accuracy does not guarantee effective teamwork; instead, failures stem from role non-adherence and miscommunication. The research suggests moving beyond simple accuracy metrics to evaluate the internal quality of agent interactions.