New AI agent evolves algorithms for math and practical applications in computing by combining the creativity of large language models with automated evaluators Large language models (LLMs) are remarkably versatile. They can summarize documents, generate code or even brainstorm new ideas. And now we’ve expanded these capabilities to target fundamental and highly complex problems in mathematics and modern computing. Today, we’re announcing AlphaEvolve, an evolutionary coding agent powered by large language models for general-purpose algorithm discovery and optimization. AlphaEvolve pairs the creative problem-solving capabilities of our Gemini models with automated evaluators that verify answers, and uses an evolutionary framework to improve upon the most promising ideas.
核心贡献 · Key contributions
AlphaEvolve 是一个由大语言模型驱动的进化式编码智能体，用于通用算法的发现与优化。 AlphaEvolve is an evolutionary coding agent powered by LLMs for general-purpose algorithm discovery and optimization.
它为 Google 的 Borg 调度器发现了一种启发式方法，恢复了全球 0.7% 的算力资源。 It discovered a heuristic for Google's Borg scheduler, recovering 0.7% of worldwide compute resources.
AlphaEvolve 优化了 TPU 设计中的一个关键算术电路，并集成到即将推出的芯片中。 AlphaEvolve optimized a key arithmetic circuit in TPU design, integrated into an upcoming chip.
它将 Gemini 的矩阵乘法内核加速了 23%，训练时间减少了 1%。 It sped up Gemini's matrix multiplication kernel by 23%, reducing training time by 1%.
AlphaEvolve 在 Transformer 模型的 FlashAttention 内核上实现了高达 32.5% 的加速。 AlphaEvolve achieved up to 32.5% speedup for FlashAttention kernel in Transformer models.
它改进了 11 维空间中的亲吻数下界，并发现了一种更快的 4x4 复数矩阵乘法算法。 It improved the kissing number lower bound in 11 dimensions and found a faster 4x4 complex matrix multiplication algorithm.
局限 · Limitations
AlphaEvolve 需要问题具有可清晰衡量的进展，这限制了其在缺乏客观指标的领域的适用性。 AlphaEvolve requires problems with clearly measurable progress, limiting its applicability to domains with objective metrics.
进化过程可能计算成本高昂，大规模搜索需要大量资源。 The evolutionary process may be computationally expensive, requiring significant resources for large-scale searches.
发现的算法可能缺乏理论保证，依赖经验验证而非形式化证明。 Discovered algorithms may lack theoretical guarantees, relying on empirical verification rather than formal proofs.
性能改进是领域特定的，可能无法泛化到其他问题实例或设置。 Performance improvements are domain-specific and may not generalize to other problem instances or settings.
系统对大语言模型的依赖意味着其创造力受限于模型的训练数据和能力。 The system's reliance on LLMs means its creativity is bounded by the models' training data and capabilities.
论文章节 · Sections（共 8）
AlphaEvolve：一个由 Gemini 驱动的编码智能体，用于设计高级算法AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms
利用大语言模型设计更好的算法Designing better algorithms with large language models
优化我们的计算生态系统Optimizing our computing ecosystem
改进数据中心调度Improving data center scheduling
辅助硬件设计Assisting in hardware design
增强 AI 训练与推理Enhancing AI training and inference
推进数学与算法发现的前沿Advancing the frontiers in mathematics and algorithm discovery