Google DeepMind researchers built AlphaEvolve, an LLM-powered system that automatically evolves game theory algorithms by rewriting their source code. Using Gemini 2.5 Pro, the system discovered new variants of Counterfactual Regret Minimization (CFR) and Policy Space Response Oracles (PSRO) that outperformed existing hand-designed algorithms in imperfect-information games like poker. The system starts with a baseline implementation, then iteratively mutates the code, evaluates candidates on proxy games, and adds successful variants to the population.
This builds directly on our previous coverage of AlphaEvolve cracking unsolved math problems through code evolution. What's striking here isn't just the performance gains—it's that the system is discovering algorithmic innovations in a field where experts have manually iterated for decades. Game theory algorithms for multi-agent scenarios have relied on human intuition and trial-and-error to develop weighting schemes and equilibrium solvers. Having an LLM systematically explore this space and find improvements suggests we're entering a phase where AI can meaningfully contribute to algorithmic research itself.
The researchers tested on established frameworks like OpenSpiel, focusing on CFR (which builds strategies by accumulating 'counterfactual regret') and PSRO (which maintains populations of policies and computes meta-strategies). They used exact best response oracles and payoff calculations, removing Monte Carlo noise from results. The fact that automated code evolution consistently found improvements over DCFR, PCFR+, and other manually-designed variants is notable—these aren't toy problems but real algorithmic challenges.
For developers building multi-agent systems or working on game-theoretic problems, this suggests LLM-driven algorithm discovery could become a viable tool. Rather than manually tweaking parameters or designing new variants, you might soon feed your algorithm to an evolution system and let it discover improvements. The question is whether this approach scales beyond game theory to other algorithmic domains—and whether the computational cost justifies the gains." "tags": ["deepmind", "algorithms", "game-theory", "code-evolution