New AI optimization framework beats Claude Code and Codex by 2.5x on the same compute budget

Researchers from Renmin University of China and Microsoft Research introduced Arbor, a framework for improving AI systems through structured, cumulative experimentation rather than ad hoc trial and error. In tests, Arbor produced more than 2.5x the verifiable gains of standard AI coding agents on real-world engineering tasks using the same resource budget.

Arbor points to a shift from one-off AI coding assistance toward autonomous optimization systems that can learn systematically from failed experiments. For enterprises, this could make complex AI deployments easier to refine in production, especially where prompts, retrieval methods, and system design choices interact in hard-to-debug ways.