A new benchmark is pushing cooperative AI agents through 100 sequential tasks — and finding that the short tests most researchers rely on miss real failure modes.
The MEAL benchmark (Multi-agent Environments for Adaptive Learning) is billed as the first purpose-built benchmark for continual multi-agent reinforcement learning. Most prior work in this area tested agents on only 3 to 10 tasks in sequence, a limit driven not by research ambition but by the slow pace of CPU-bound simulation. MEAL sidesteps that by running on JAX with GPU acceleration, compressing a 100-task training sequence down to a few hours on a single GPU. The researchers say failure modes that simply do not appear in shorter sequences emerge clearly at that scale.
This matters because "lifelong learning" — the idea that an AI system should accumulate knowledge across tasks without forgetting earlier ones — has been a stated goal of RL research for years. A benchmark capped at 10 tasks is a poor test of that goal. MEAL gives the field a shared, reproducible way to find out whether proposed solutions actually hold up over time, rather than just over a brief sprint.
Cooperative multi-agent settings add another layer: agents must adapt not just to new tasks but to the shifting behavior of teammates, a dynamic that single-agent benchmarks ignore entirely. Whether the research community adopts MEAL as a standard, or treats it as one of several competing yardsticks, will depend on whether the GPU requirement is a feature or a barrier for labs without deep compute budgets.