Bench-MFG: A Benchmark Suite for Learning in Stationary Mean Field Games

Researchers in the field of Mean Field Games and Reinforcement Learning have introduced Bench-MFG, a comprehensive benchmark suite designed to evaluate algorithms for stationary Mean Field Games, addressing the lack of standardized evaluation protocols in the field, as detailed in their paper published on arXiv on February 12, 2026, which aims to solve the current fragmentation in research assessment. The Bench-MFG project emerges from the growing intersection of Mean Field Games (MFGs) and Reinforcement Learning (RL), which has developed numerous algorithms for solving complex large-scale multi-agent systems. However, the research community has faced significant challenges due to the absence of standardized evaluation protocols, forcing researchers to rely on customized, isolated, and often overly simplistic testing environments. This lack of standardization has created substantial difficulties in accurately assessing the robustness, generalization capabilities, and potential failure modes of emerging methodologies in the field. The introduction of Bench-MFG represents a significant step toward establishing more rigorous evaluation standards in the MFG and RL research communities. By providing a standardized suite of benchmarks, researchers can now more effectively compare different approaches, identify strengths and weaknesses, and drive innovation in algorithm development. The benchmark suite is expected to facilitate more reproducible research and accelerate progress in solving increasingly complex multi-agent systems through standardized testing protocols and comprehensive evaluation metrics.