AI4AI Review Proposes L0–L3 Research Control Framework
A review by teams from Nankai University and Zhongguancun Academy divides AI-driven research into three decisions: search direction, search space, and evaluation signals. Its L0–L3 framework describes who controls each decision, while four branches cover AI4AI learning, inference, data, and infrastructure. The review also argues that independent validation is needed when AI improves both programs and tests.
The framework provides a concrete way to distinguish AI-assisted execution from AI control over research priorities, candidate designs, and success criteria.
新智元 2026-11-14 17:06 北京
新智元报道
你让AI自动设计一种新的推理算法:它能读论文、写代码、跑实验,还能根据结果不断修改方案。
看起来,整个研究过程已经交给了AI。
但如果研究问题由人类提出、允许修改的范围由人类划定、判断成功的标准也由人类设定,那么AI真正掌握了多少研究决策?它是在替人做实验,还是已经开始决定如何做研究?
#Nankai University#Zhongguancun Academy#FunSearch#DreamCoder#Eureka#AutoML-Zero#DSPy#DoReMi#AlphaEvolve