【Proposal】Systematic Idea Refinement for Machine Learning Research Agents
Keywords: Automatic Machine Learning, Automatic Research, Research Idea Generation, LLM, Agent
TL;DR: We propose to automatically generate and refine research idea as guidelines for ML research agents via a plug-and-play multi-agent method.
Abstract: This project aims to enhance the machine learning (ML) research capabilities of large language model (LLM)-empowered agents beyond basic code generation. While recent advancements have demonstrated the use of single-agent systems for generating code across various ML tasks, these methods often focus solely on improving code validity. They lack the ability to explore diverse methodologies for a given problem, which limits their adaptability and performance. To address this gap, the proposed project develops a multi-agent framework that systematically refines research idea guidelines through automatic proposal, feedback integration, and inference-time scaling. By incorporating multi-level feedback from LLM judgments, code generation processes, and experimental results, this approach enables agents to explore a broader range of solution pathways, similar to human researchers. The framework is plug-and-play on code generation agents, and will be evaluated on $\textbf{a total number of 75 Kaggle competitions}$. The expected outcome is an improvement in the understanding and performance of machine learning research agents through a comprehensive exploration of methodological ideas.
Submission Number: 18
Loading