Efficient Uncoupled Learning Dynamics with $\tilde{O}\!\left(T^{-1/4}\right)$ Last-Iterate Convergence in Bilinear Saddle-Point Problems over Convex Sets under Bandit Feedback

A research team from the University of Washington and Georgia Tech has published a significant advance in algorithmic game theory and multi-agent learning. Their paper, accepted at AISTATS 2026, presents a new 'uncoupled' learning algorithm that solves bilinear saddle-point problems—a mathematical framework that models competition between two agents, like in generative adversarial networks (GANs) or strategic pricing. The key breakthrough is achieving 'last-iterate convergence,' meaning the algorithm's daily outputs steadily approach an optimal equilibrium, rather than just averaging out over time. This is crucial for real-world deployment where you need reliable performance at every step, not just on average.

The algorithm operates under the challenging 'bandit feedback' setting, where each player only observes the payoff from their own chosen action, not the full structure of the game. By cleverly integrating techniques from experimental design with a tailored Follow-The-Regularized-Leader (FTRL) approach, the team proved a convergence rate of O(T^{-1/4}), where T is the number of rounds. Importantly, it only requires an efficient linear optimization oracle over the players' action sets, making it practical for complex domains. This work provides a foundational tool for building more stable and predictable multi-agent AI systems, from training GANs to developing robust economic and strategic algorithms where agents learn through pure interaction.