Arbor AI Optimization Framework Beats Claude Code and Codex by Over 2.5×

Arbor’s researchers at Renmin University and Microsoft Research present a breakthrough

AI Optimization Framework: How Arbor Redefines Autonomous Coding

that turns the chaotic trial‑and‑error of production AI agents into a disciplined learning loop. In early trials, the system achieved ↑ 2.5x the verified gains of Claude Code and Codex while staying within identical compute budgets.

“Automation can keep an AI working for a very long time — but a loop is not the same as progress,”

says Jiajie Jin, co‑author, in an interview with Reuters. The architecture splits responsibilities between a long‑lived coordinator, which curates a hypothesis tree, and short‑lived executors that test individual ideas in isolated git worktrees. This separation prevents entangled changes and provides clean attribution for each lever—chunking, prompting, retrieval—allowing engineers to pinpoint the exact source of a performance jump. On the BrowseComp benchmark, Arbor lifted held‑out accuracy from ↑ 45.33% to 67.67%, while Claude Code plateaued near 50% and Codex lingered at 53.33%. The framework also resisted reward hacking; in Terminal‑Bench 2.0, its development score lagged behind Claude Code but its held‑out score topped at 77.36, confirming real‑world transfer. Arbor integrates seamlessly with existing Git workflows: its output is a regular branch that can pass through standard code review, CI pipelines, and human vetting. The primary cost is token consumption for the coordinator and compute for parallel worktrees, making it best suited for tasks with reliable metrics and ample time horizons, such as pipeline tuning or model‑training recipe optimization. Future iterations aim to expand the hypothesis node to carry multi‑dimensional metrics—accuracy, latency, cost—enabling Pareto‑optimal searches. For enterprises eager to automate continuous improvement without sacrificing traceability, Arbor offers a disciplined, scalable path forward.