Learn2Match uses MARL to handle dynamic two-sided matching markets

Traditional two-sided matching models (e.g., for job markets or dating platforms) assume immediate feedback about fixed preferences, ignoring how information unfolds over time through interviews, repeated interactions, and relationship dissolution. To address this gap, Haijing Zong and colleagues introduce a framework that formulates two-sided matching as a partially observable Markov game. Their model incorporates costly pre-match screening, noisy post-match observations, evolving latent user profiles, and the option to end or continue a match. They instantiate this in Learn2Match, a multi-agent reinforcement learning benchmark designed for dynamic matching markets.

In experiments, they compare independent PPO (a popular MARL algorithm) against CA-ETC, a bandit-style baseline. PPO achieves higher cumulative social welfare and lower cumulative regret under temporally extended feedback. However, PPO incurs higher information-friction loss—a metric measuring welfare loss from incomplete preference revelation. This reveals that end-to-end MARL lacks the coordinated exploration structure of matching-bandit methods. Learn2Match is positioned as a benchmark to develop next-generation matching algorithms that are adaptive like RL, statistically disciplined like bandits, and structurally aware like stable-matching mechanisms, potentially transforming how we build real-world matching platforms.