Core Insights - The article discusses the limitations of existing large language models in machine learning engineering, particularly in tasks like AutoML and Kaggle competitions, where continuous iteration and high-performance tuning are essential [1][2] - AutoMLGen, developed by Shanghai Artificial Intelligence Laboratory and East China Normal University, is introduced as a new intelligent programming framework that integrates general large model reasoning with domain knowledge [1][2] Group 1: AutoMLGen Framework - AutoMLGen is designed to enhance the capabilities of large language models beyond code generation, enabling continuous optimization and experience reuse [4][6] - The framework consists of three main modules: a knowledge base, Monte Carlo Graph Search (MCGS), and a fine-grained operator library, which together create a self-evolving loop from experience guidance to intelligent exploration and solution refinement [6][8] Group 2: Knowledge Base - The knowledge base in AutoMLGen systematizes the experience of skilled machine learning engineers, covering model selection, feature processing, and strategy design [7] - During the task initiation phase, AutoMLGen autonomously decides whether to utilize domain knowledge, effectively alleviating the cold start problem while maintaining the independence of the intelligent agent's decisions [7] Group 3: Monte Carlo Graph Search (MCGS) - MCGS innovatively introduces a graph structure to the search process, allowing for dynamic fusion and sharing of nodes and trajectories across different branches, thus enhancing efficiency in complex tasks [8] - Four core mechanisms drive the continuous evolution of the intelligent agent: main expansion, intra-branch evolution, cross-branch reference, and multi-branch aggregation [8] Group 4: Fine-Grained Operator Library - The fine-grained operator library in AutoMLGen defines the evolution methods between different solutions, facilitating a coherent and efficient optimization process [9] - This mechanism allows the intelligent agent to transition from a code generator to an AI engineer capable of proactive reflection and improvement [9] Group 5: Performance Results - AutoMLGen achieved a 36.4% average medal rate and an 18.7% gold medal rate on the MLE-Bench leaderboard, outperforming existing systems while using only half the standard computation time (12 hours) [12][19] - In the MLE-Bench-Lite tests, AutoMLGen maintained a significant lead, demonstrating consistent performance and excellent generalization capabilities [12] Group 6: Future Prospects - The emergence of AutoMLGen signifies a shift in the capabilities of intelligent agents in complex engineering and algorithm design tasks, showcasing AI's potential for autonomous exploration and continuous improvement [19][20] - The framework's principles are expected to extend to broader intelligent system paradigms, paving the way for future developments in AI that can actively understand, improve, and innovate solutions [20]

InternAgent 团队 投稿 量子位 | 公众号 QbitAI 在代码层面，大语言模型已经能够写出正确而优雅的程序。但在机器学习工程场景中，它离真正"打赢比赛"仍有不小差距。 因为像AutoML任务与Kaggle竞赛，不仅要求生成可运行的代码，更要求在数据处理、算法设计层面持续迭代与高性能调优。过去，这一过程 往往依赖专家经验与反复试错，使模型难以高效突破瓶颈。 然而，现有基于大模型的机器学习智能体仍受限于两大问题： 简单来说，就是它们会写代码，却还不会"聪明地优化"代码。 在此背景下，上海人工智能实验室联合华东师范大学提出了 AutoMLGen ，一个融合通用大模型推理与领域知识的智能编程框架。 其核心为自研的 蒙特卡洛图搜索（MCGS） ，通过"分支—节点动态融合"打破传统MCTS的孤立局限，让不同搜索分支可共享高价值节点； 并结合 领域知识库 与 算子级优化 ，将搜索重点快速聚焦到有效空间，实现轨迹复用、跨分支聚合与过程学习。 AutoMLGen在仅使用DeepSeek-R1模型的情况下，以36.4%的平均奖牌率和18.7%的金牌率登顶MLE-Bench榜单 ，用标准时长一半（12 小时）的计算预 ...