New AI Framework Outperforms Claude, Codex by 2.5x

Researchers at Renmin University of China and Microsoft Research have developed Arbor, a framework that transforms AI optimization from trial-and-error guessing into a structured learning process. It organizes hypotheses, experiments, and insights into a tree structure, enabling the system to learn from past failures. In practical tests, Arbor outperformed Claude Code and Codex by 2.5x on the same compute budget. The framework addresses a common production problem where AI agents hallucinate or miss constraints, making it difficult to pinpoint which specific adjustments actually fix the issue.