MARS: Modular Agent with Reflective Search for Automated AI Research
Jiefeng Chen, Bhavana Dalvi Mishra, Jaehyun Nam, Rui Meng, Tomas Pfister, Jinsung Yoon · 2026
Read MARS: Modular Agent with Reflective Search for Automated AI Research by Jiefeng Chen, Bhavana Dalvi Mishra, Jaehyun Nam, Rui Meng, Tomas Pfister, Jinsung Yoon (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-02660.txt.
Description
Automating AI research itself is the goal of a 2026 system from Jiefeng Chen and 5 co-authors. Machine learning engineering differs from ordinary coding: evaluation means expensive model training, and knowing which change caused a gain is opaque. Language-model agents tend to write monolithic scripts that ignore both problems.
Their framework, MARS, rests on three pillars. Budget-aware planning searches with Monte Carlo tree search under explicit cost constraints. Modular construction follows a design-decompose-implement pipeline suited to complex research repositories. Comparative reflective memory tackles credit assignment by analyzing differences between solutions and distilling high-signal lessons. On MLE-Bench the system reached the best performance among open-source frameworks and stayed competitive with top leaderboard methods. Most lessons the agent actually used came from transferring insights across different search branches.
Catalog record
| CREATOR | Jiefeng Chen, Bhavana Dalvi Mishra, Jaehyun Nam, Rui Meng, Tomas Pfister, Jinsung Yoon |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 130 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-02660