MARL methods score 0% on new composite task challenge benchmark

The Composite Task Challenge (CTC) is a new benchmark suite proposed by Yurui Li and colleagues at Zhejiang University to address a glaring blind spot in cooperative multi-agent reinforcement learning (MARL). While many real-world applications — from warehouse robotics to autonomous vehicle coordination — rely on efficient division of labor (DOL), existing MARL benchmarks rarely enforce DOL as a strict requirement. CTC flips that: its tasks are built on two core design principles. First, DOL is a necessary condition for success; agents must specialize and coordinate their roles. Second, failure in any atomic subtask immediately fails the entire task, making cooperation non-negotiable. This creates a stress test for methods that claim to handle complex multi-agent coordination.

The results are stark. Researchers evaluated nine representative cooperative MARL algorithms — including QMIX, MAPPO, and VDN — across all CTC tasks. Every single method achieved a zero percent test winning rate, indicating that current approaches completely break down when forced to both divide labor and avoid any single-point failure. To show the tasks are solvable, the team also introduced a guiding solution that managed non-zero test winning rates on all tasks, but its performance remained far from optimal. This benchmark exposes a fundamental limitation in today's MARL methods and provides a rigorous testbed for future research aimed at building AI agents that can truly cooperate in dynamic, interdependent environments.