Google DeepMind's AlphaEvolve Uses Gemini 2.5 Pro to Rewrite Its Own Algorithms, Beats Human-Designed Game Theory Code
Google DeepMind has built a system called AlphaEvolve that uses Gemini 2.5 Pro to do something genuinely unprecedented: it writes its own algorithms. Not just generates code snippets or suggests optimizations, but actually evolves complete algorithms through iterative mutation and selection, and the results are outperforming decades of human-designed approaches in game theory—a domain where human mathematicians have been refining techniques for years. AlphaEvolve recently discovered VAD-CFR, a new variant of Counterfactual Regret Minimization for imperfect-information games, which dynamically adjusts its discounting based on measured volatility—a mechanism that never occurred to human researchers working in the field. It matches or beats established baselines like DCFR and PCFR+ across multiple test games.
The implications reach well beyond games. Multi-agent reinforcement learning systems underpin automated trading, autonomous negotiation platforms, and cybersecurity defense frameworks—all areas where the quality of the underlying algorithm directly translates to real-world advantage. When a system can autonomously explore the space of possible algorithmic variations at scale and select for performance, the traditional bottleneck of human mathematical creativity gets removed. This doesn't make human researchers obsolete, but it changes the nature of their role: less generation, more evaluation and direction.
This is the clearest signal yet that LLMs can do more than pattern-match their training data—they can discover genuinely novel solutions in structured mathematical domains. The game theory result is particularly striking because these algorithms are verifiable: you can prove whether VAD-CFR is better or worse, and it is demonstrably better in multiple cases. That's not a benchmark improvement, it's a capability shift. The question of what else in complex software engineering might be similarly automatable through evolutionary search with an LLM at the core is now wide open.