ALPHAEVOLVE DISCOVERS NEW MULTI-AGENT LEARNING ALGORITHMS

Summary from the TLDR AI newsletter

Using an evolutionary coding agent powered by large language models, the work automatically generated new algorithms for imperfect-information games across CFR and PSRO paradigms. The discovered VAD-CFR and SHOR-PSRO variants introduced adaptive discounting and hybrid optimistic meta-solvers, showing improved empirical convergence over strong baselines.