SWE-Search introduces a framework that enables software agents to strategically explore solution paths, improving their ability to solve complex software engineering tasks.

• Combines Monte Carlo Tree Search (MCTS) with LLMs to explore multiple solution paths and avoid dead ends
• Implements iterative refinement allowing agents to learn from failures and revise approaches
• Achieves state-of-the-art results on the SWE-bench repository-level coding benchmark
• Enables agents to backtrack when solutions prove ineffective, mimicking human problem-solving

This research significantly advances AI coding assistants by moving beyond linear solution approaches, making them more reliable for real-world software engineering challenges.

SWE-Search: Enhancing Software Agents with Monte Carlo Tree Search and Iterative Refinement