[Paper Review] MARS: Modular Agent with Reflective Search for Automated AI Research

Jiefeng Chen, Bhavana Dalvi Mishra+4 more|arXiv (Cornell University)|Feb 2, 2026

Scientific Computing and Data Management0 citations

TL;DR

MARS introduces Budget-Aware MCTS, Modular Decomposition, and Comparative Reflective Memory to automate AI research, achieving state-of-the-art open-source performance on MLE-Bench and strong cross-branch generalization.

ABSTRACT

Automating AI research differs from general software engineering due to computationally expensive evaluation (e.g., model training) and opaque performance attribution. Current LLM-based agents struggle here, often generating monolithic scripts that ignore execution costs and causal factors. We introduce MARS (Modular Agent with Reflective Search), a framework optimized for autonomous AI research. MARS relies on three pillars: (1) Budget-Aware Planning via cost-constrained Monte Carlo Tree Search (MCTS) to explicitly balance performance with execution expense; (2) Modular Construction, employing a "Design-Decompose-Implement" pipeline to manage complex research repositories; and (3) Comparative Reflective Memory, which addresses credit assignment by analyzing solution differences to distill high-signal insights. MARS achieves state-of-the-art performance among open-source frameworks on MLE-Bench under comparable settings, maintaining competitiveness with the global leaderboard's top methods. Furthermore, the system exhibits qualitative "Aha!" moments, where 63% of all utilized lessons originate from cross-branch transfer, demonstrating that the agent effectively generalizes insights across search paths.

Motivation & Objective

• Motivate and address the distinct challenges of automating AI research, including expensive evaluations and opaque credit attribution.
• Propose a framework (MARS) that balances performance with computational cost through budget-aware planning.
• Promote modular repository-level construction to manage architectural complexity and enhance testability.
• Introduce Comparative Reflective Memory to distill causal insights and guide long-horizon exploration.

Proposed method

• Implement Budget-Aware Monte Carlo Tree Search (MCTS) to balance performance with execution cost via an efficiency-guided reward (Eq. 4).
• Adopt a Modular Design-Decompose-Implement pipeline to replace monolithic scripts with independent, testable modules and enable Diff-Based Editing for atomic updates.
• Introduce Comparative Reflective Memory to extract high-signal lessons by comparing current solutions to the best-known ones, including structured debugging and solution lessons.
• Use a three-part framework (Task Preparation, Resource-Aware Planning, Modular Decomposition, Reflective Memory) to convert long-horizon AI research into a repository-level problem.
• Evaluate on MLE-Bench under a 24-hour wall-clock budget, reporting Above Median, Bronze, Silver, Gold, and Any Medal metrics, with ablations validating each component.

Figure 1 : The “Aha!” moment of MARS on the challenging iMet-2020-FGVC7 task. The visualization tracks validation performance gains triggered by specific strategic lessons. While existing methods fail to reach medal-level performance, MARS progressively refines its strategy – evolving from a lightwe

Experimental results

Research questions

• RQ1 How can budget-aware planning improve efficiency in long-horizon AI research tasks?
• RQ2 Does modular decomposition improve solution quality and maintainability for complex research pipelines?
• RQ3 Can comparative reflective memory enable effective credit assignment and faster long-horizon learning?
• RQ4 What is the impact of lesson learning on cross-branch transfer and exploration dynamics?
• RQ5 How does MARS perform relative to open-source baselines on MLE-Bench under realistic constraints?

Key findings

• MARS achieves state-of-the-art performance among open-source frameworks on MLE-Bench under comparable settings.
• MARS+ with increased compute achieves the highest Above Median, Gold Medal, and Any Medal rates, outperforming leading baselines.
• Ablation studies show that both Modular Decomposition and Lesson Learning significantly boost performance.
• Budget-Aware MCTS yields higher effective solution rates and prefers faster candidates when performance is comparable, accelerating discovery.
• Lessons exhibit high utilization and cross-branch transfer, indicating effective generalization of insights across search paths.

Figure 2 : Overview of the MARS Framework. MARS reformulates long-horizon coding as a search for an optimal software repository. (1) Task Preparation: The agent grounds the abstract problem (Instruction, Environment, Objective) tuple by exploratory analysis of the given dataset and metadata. (2) The