Nankai Team Proposes L0-L3 Framework to Dissect AI Autonomy in AI Research
新智元 · wechat · 2026-11-14
A survey from Nankai University and Zhongguancun Academy breaks AI-improving-AI (AI4AI) research into three key decisions — search direction, search space, and evaluation signal — and proposes an L0-L3 control framework.
- L0: humans set all three decisions; L1: AI picks what to search next (e.g. FunSearch); L2: AI expands the searchable space (e.g. DreamCoder's reusable abstractions); L3: AI shapes evaluation signals (e.g. Eureka's reward functions).
The survey also organizes AI4AI into four branches — Learning, Inference, Data, and Infrastructure (citing AlphaEvolve's matmul and TPU circuit optimizations) — and argues that compressing search spaces can matter as much as expanding them. When AI builds its own evaluators, independent verification paths (hidden tests, delayed real-world results) are needed to confirm genuine progress, and recursive self-improvement requires cross-generation evidence: multi-generation experiments, fixed-budget baselines, and transfer to unseen tasks.
More from AGI Musings
- AGI endgame: either a balanced multipolar world or a centralized panopticon — CedricMakes · 2026-11-14
- "60% of people around you have some form of AI psychosis," says Alsop III — StewartalsopIII · 2026-11-14
- AI won't make everyone equal—it will split the smart from the rest worse than money — prasenx · 2026-11-14
- Toddlers learn new verbs only from responsive partners; a call to rerun the test with AI — AnnaCiaunica · 2026-11-14
- Developer muses: in Murphy's world, billions of interdependent events make rare failures daily — bingxu_ · 2026-11-14
- Solo dev: agents made me a hundred-decisions-a-day generalist — and unhireable — rudrank · 2026-11-14