Regret matching and counterfactual regret minimization are workhorse algorithms for solving normal-form and extensive-form games, but they have historically lacked strong theoretical convergence guarantees. A new paper on arXiv introduces IR-PRM+, a variant that incorporates previous-utility prediction to close this gap.

The authors report that IR-PRM+ achieves constant regret, meaning the cumulative regret stays bounded rather than growing with the number of iterations. They also show best-iterate convergence, where the strategy produced at each step improves over time, rather than only the average of past strategies. These are stronger properties than what is typically proven for regret-matching algorithms.

Because the paper is a single source, there are no conflicting results to compare. The claims are theoretical and presented in the abstract; the full proof would be in the body of the paper. If verified, this could make regret-based solvers more reliable for game-theoretic applications that depend on the latest iterate rather than an averaged policy.