Researchers from MIT, the University of Texas at Austin, the University of California Berkeley, Carnegie Mellon University, and New York University have demonstrated that policy gradient machine learning algorithms outperform specialized game-theoretic approaches in multi-agent strategic environments. The findings were presented in April at the International Conference on Learning Representations in Rio de Janeiro. The study challenges a longstanding assumption in artificial intelligence and game theory. For decades, the field has prioritized specialized game-theoretic algorithms for training neural networks in imperfect-information games, where participants must make decisions without complete knowledge of their opponents actions. Policy gradient methods, which optimize decision-making strategies through iterative feedback loops, were originally designed for single-agent reinforcement learning and were largely dismissed as inadequate for complex, competitive environments. The new research provides a rigorous, even-handed framework to test these competing approaches, shifting focus from developing novel algorithms to establishing a standardized benchmark for performance evaluation. Central to the study is the metric of exploitability, which quantifies how effectively a player performs against a worst-case adversary. A score of zero indicates optimal play, while higher values reflect suboptimal decision-making. The research team developed a testing environment capable of analyzing games with up to 30 billion possible states, a significant increase in scale compared to prior studies that typically examined environments with fewer than 300,000 states. Experiments spanned five distinct games, including two variants of Phantom Tic-Tac-Toe, two imperfect-information versions of Hex, and Liar’s Dice. Across all tested environments, neural networks trained with policy gradient methods achieved substantially lower exploitability scores than those trained with traditional game-theoretic algorithms. In direct head-to-head competitions, the policy gradient models consistently defeated their game-theory counterparts. The research team emphasized that their primary contribution is not a new algorithm, but a standardized, open-source benchmarking platform. The software integrates seamlessly with the widely used OpenSpiel library and requires only a single line of code adjustment, allowing it to run on standard consumer hardware without supercomputing resources. Experts note that the implications extend well beyond recreational gaming. The researchers and independent reviewers, including experts from Google DeepMind, argue that the principles governing hidden-information environments apply directly to high-stakes domains such as financial trading, international negotiations, and military strategy. By validating policy gradient methods in complex strategic settings, the study suggests that modernizing classical machine learning techniques remains a highly effective pathway for solving real-world multi-agent problems. The open-source release of the benchmarking framework is expected to accelerate standardized evaluations across the artificial intelligence community, enabling more reliable comparisons and faster advancements in strategic machine learning.