Chinese and Microsoft Research Team Develops AI Optimization Framework 「Arbor」

A research team from Renmin University of China and Microsoft Research has developed a new framework 「Arbor」 to support autonomous improvement of AI systems. The framework is designed to realize more efficiently a technique called 「Autonomous Optimization (AO)」, in which AI agents learn by repeatedly engaging in trial-and-error. In validation using actual engineering tasks, Arbor achieved verifiable performance improvements of 2.5 times or more compared to standard AI coding agents within the same computational resource budget.

Autonomous optimization refers to a process in which an AI agent is given an 「improvement target」 such as machine learning code or a data processing pipeline, and repeatedly conducts experiments and incorporates feedback to continuously refine itself without human direction at every step. This approach has drawn attention for its potential to automate continuous improvement of software, but existing agents had significant structural flaws. Because each trial is processed independently, insights gained from past experiments are not carried forward to the next trial. Jiajie Jin, a co-author of the paper, stated: 「We can keep AI running for extended periods through automation, but a continuing loop does not necessarily mean progress.」

Arbor solves this challenge by managing hypotheses, experiments, and obtained insights in a 「tree structure」. A tree structure is a way of organizing data that hierarchically arranges information through branching, allowing systematic retention of records of past failures and successes. This enables agents to devise the next improvement strategy while referencing previous attempts, avoiding repeated failures while allowing learning to accumulate. While previous agents had a structure that 「processed each trial separately」, Arbor is novel in that it systematically provides a structure capable of accumulating experience for the first time.

One scene where this mechanism becomes particularly important is the operation of enterprise AI systems. For example, when developing and operating an AI agent that searches internal documents to answer employee questions, problems often occur in production environments where the system outputs incorrect information or ignores important constraints, even though it functions well in the development phase. To fix these problems, it becomes necessary to simultaneously adjust the document segmentation method, search logic, and system instructions, making it difficult to identify which changes contributed to improvements. Arbor's tree structure-based management makes it easier to track the effects of such complex adjustments.

The research team explains that Arbor can be directly applied to automating continuous improvement of enterprise AI systems. In current AI development, system performance improvements often depend on manual trial-and-error by engineers. If frameworks like Arbor become practical, it would be possible to automate part of that process. However, what is currently being released represents research-stage results, and details regarding actual implementation in products and services remain unclear at this time.