MARS: Modular Agent with Reflective Search for Automated AI Research скачать в хорошем качестве

MARS: Modular Agent with Reflective Search for Automated AI Research

Introducing MARS, an agent framework that autonomously performs complex artificial intelligence research and machine learning engineering tasks. Unlike traditional systems that generate a huge single code and overlook resource efficiency, MARS has adopted a repository-level modular structure to increase code maintainability and accuracy. In particular, we explore efficient solutions by introducing a budget-aware Montecarlo tree search (MCTS) algorithm that balances performance and computing costs. In addition, through the reflectional memory mechanism, we accumulate lessons learned from past successes and failures and strategically use them for subsequent work. As a result of the experiment, the system overwhelmed existing agents in high-level benchmarks such as MLE-Bench, demonstrating strategic insight similar to that of human engineers. As a result, this source emphasizes the importance of structural design and resource optimization for autonomous driving research agents. https://arxiv.org/pdf/2602.02660