New AI Framework Beats Top Coding Agents by 2.5x

Arbor, a new AI optimization framework, achieves more than 2.5 times the verifiable performance gains of Claude Code and Codex on the same compute budget.

Researchers from Renmin University of China and Microsoft Research have developed a new AI framework called Arbor. It helps AI agents learn from past mistakes and make smarter improvements over time. In tests, Arbor delivered more than 2.5 times the verifiable performance gains of standard AI coding agents like Claude Code and Codex.

Arbor organizes the research process into a tree structure. A coordinator agent manages the overall strategy and creates hypotheses. Short-lived executor agents test each hypothesis in isolated environments. This design allows the system to explore multiple ideas at the same time without mixing up results.

The framework also prevents reward hacking by using a strict merge gate. Before any improvement is accepted, Arbor tests it against a separate evaluation set. This ensures that the gains are real and transfer to real-world applications. Arbor has shown strong results on tasks like optimizing search agents and machine learning pipelines.