Google DeepMind’s research lets an LLM rewrite its own game theory algorithms — and it outperformed the experts

Google DeepMind’s AlphaEvolve used Gemini 2.5 Pro to mutate algorithm source code directly, searching for new multi-agent reinforcement learning methods.

Applied to CFR and PSRO in OpenSpiel’s imperfect-information games, it used exact best responses and payoff calculations on poker-like benchmarks.

The system discovered new variants such as Volatility-Adaptive Discounted CFR and also improved PSRO meta-strategy solvers.

On held-out games including 3-player Kuhn Poker, 2-player Leduc Poker, Goofspiel, and Liars Dice, it matched or beat standard expert-designed methods.

The work suggests LLMs can automate game-theory and RL algorithm design, reducing manual trial-and-error and surfacing stronger strategies.