ML-Master - An AI expert agent launched by Shanghai Jiao Tong University

What is ML-Master?

ML-Master is an AI expert agent developed by the Agents team at the School of Artificial Intelligence, Shanghai Jiao Tong University. It performed exceptionally well in OpenAI's authoritative benchmark test, MLE-bench, ranking first with an average medal rate of 29.3%, surpassing competitors such as Microsoft's RD-Agent and OpenAI's AIDE. ML-Master employs an innovative paradigm of "deep integration of exploration and reasoning," simulating the cognitive strategies of human experts and integrating extensive exploration with deep reasoning, significantly improving AI performance in machine learning engineering. It utilizes a balanced multi-path exploration and controllable reasoning modules, achieving efficient collaboration between the two through an adaptive memory mechanism.

Main functions of ML-Master

• Deep integration of exploration and reasoningML-Master significantly improves AI performance by simulating the cognitive strategies of human experts through an innovative "deep fusion of exploration and reasoning" paradigm, integrating extensive exploration and deep reasoning.
• Excellent performance:
  • In OpenAI's MLE-bench benchmark test, ML-Master ranked first with an average medal rate of 29.3%, significantly ahead of Microsoft's RD-Agent (22.4%) and OpenAI's AIDE (16.9%).
  • 93.3% of the tasks submitted valid solutions, and 44.9% of the tasks surpassed those of half of the human participants, demonstrating strong generalization ability and stability.
  • It is extremely computationally efficient, completing the test in just 12 hours, with a computational cost that is only half that of the baseline method.
• Strong self-evolutionary abilityML-Master continuously improves the quality of the solution during multiple rounds of task execution, ultimately achieving a performance improvement of over 120% compared to the initial version.

ML-Master's technical principles

• Balanced Multi-trajectory Exploration
  • MCTS-inspired tree searchThe AI R&D process is modeled as a decision tree, with each node representing the state of an AI solution.
  • Parallel exploration strategySimultaneously explore multiple solution branches, breaking through the limitations of traditional serial exploration and significantly improving exploration efficiency.
  • Dynamic priority adjustment: Dynamically allocate computing resources based on the potential value of each branch to avoid ineffective exploration.
• Steerable Reasoning
  • Adaptive memory mechanismIt accurately extracts key information, avoids information overload, and intelligently filters effective information from historical explorations to ensure that the reasoning process is based on more relevant knowledge.
  • Contextualized decision making: Conduct evidence-based analysis by combining specific implementation feedback and success stories to avoid making decisions based on guesswork.
  • Closed-loop learning systemThe results of the exploration are fed back into the reasoning process in real time, forming a virtuous cycle of "exploration → reasoning → optimization → re-exploration".
• Adaptive Memory
  • Intelligent memory constructionThe exploration module automatically collects execution results, code snippets, and performance metrics, while selectively integrating key information from parent nodes and parallel sibling nodes.
  • Embedded reasoning decisionThe memory information is directly embedded into the decision-making part of the reasoning model, ensuring that each reasoning is based on specific historical execution feedback and experience from diverse explorations.
  • Co-evolution mechanismThe results of reasoning guide the direction of subsequent exploration, and the exploration experience continuously enriches the reasoning process, achieving a deep integration of exploration and reasoning.

ML-Master project address

• Project official websitehttps://sjtu-sai-agents.github.io/ML-Master/
• Github repository: https://github.com/sjtu-sai-agents/ML-Master
• arXiv technical paperhttps://arxiv.org/pdf/2506.16499

Application scenarios of ML-Master

• Machine learning task automationML-Master, through its "deep fusion of exploration and inference" technical framework, can automatically complete the entire machine learning process from model training and data preparation to experimental execution. It performs exceptionally well in OpenAI's MLE-bench benchmark test, demonstrating its efficiency and accuracy in handling complex machine learning tasks.
• AI development efficiency improvementML-Master significantly improves the efficiency of AI development by balancing multi-path exploration and controlled inference modules. It is suitable for AI projects that require rapid iteration and optimization.
• AI Self-Evolution and OptimizationML-Master possesses strong self-evolution capabilities, continuously improving solution quality through multiple rounds of task execution. It is suitable for AI systems requiring long-term optimization and self-improvement, such as adaptive learning and optimization tasks in complex environments.
• Multi-domain task coverageML-Master can be extended to other fields that require AI-driven autonomous optimization, such as materials science, medical diagnostics, and financial trading. For example, the framework can be used for material property prediction, new material discovery, and production process optimization.
• Sentiment Analysis and Text ProcessingThe technical principles of ML-Master can also be applied to the field of natural language processing, such as sentiment analysis and opinion mining. It can perform semantic representation of text, and based on this, perform sentiment classification and opinion extraction, making it suitable for scenarios such as consumer decision-making and public opinion analysis.