Skip to content
WeChat · AI Era (Xinzhiyuan)·Indexed by this site · Originally published· Alignment →· Its place in the scan record →

Nankai University and Zhongguancun Academy team publish "Let AI Research AI" survey: three key decision frameworks re-examine research autonomy

「让AI研究AI」最新综述:三项关键决策框架,重新审视研究自主性

AI summary

A team from Nankai University and Zhongguancun Academy published a survey on letting AI research AI (AI4AI), breaking down AI's decisions in research into three items: search direction, search space, and evaluation signal, and proposing an L0 to L3 control-source framework: at L0, humans or external rules give the three decisions; at L1, as in FunSearch, AI decides the next search step; at L2, as in DreamCoder, AI expands the searchable solutions; at L3, as in Eureka, AI generates reward functions that affect training. The survey also divides AI4AI by improvement target into four branches—learning, reasoning, data, and infrastructure—and lists applications such as AlphaEvolve optimizing TPU arithmetic circuits. The authors emphasize that once AI begins constructing tests or reward functions, internal scores alone cannot distinguish improved program capability from tests becoming easier, and external validation paths such as held-out tests and independent experiments must be retained; recursive self-improvement requires cross-generation evidence such as multi-generation experiments, equal-budget baselines, and transfer to unseen tasks. Paper address: https://www.preprints.org/manuscript/202610.2999

Read originalmp.weixin.qq.com