MARS: Modular Agent with Reflective Search for Automated AI Research

MARS (Modular Agent with Reflective Search) is a framework designed to automate AI research by handling complex machine learning engineering (MLE) tasks. Unlike general software engineering, MLE involves computationally expensive evaluations like model training and opaque performance attribution, which current LLM-based agents struggle with. MARS addresses these challenges through three pillars: Budget-Aware Planning using cost-constrained Monte Carlo Tree Search (MCTS) to balance performance and execution cost; Modular Construction via a 'Design-Decompose-Implement' pipeline for managing complex research repositories; and Comparative Reflective Memory that analyzes solution differences to improve credit assignment. The framework achieves state-of-the-art results in autonomous AI research, as detailed in the arXiv paper 2602.02660.