Survey overview: component benchmarks and environments on the left, six task pressures and the model-harness-environment-evaluation stack in the middle, model-side and harness-side methods on the right.

Blog · AI4AI explained · September 2026

What is AI4AI? Key findings of the AI4AI survey

A short English read-through of the AI4AI survey: what the field is, what already works, and what is still missing.

Kai Wu^1,*, Hao Lyu^1,*, Zhen Luo^1,*, Chaofan Wang², Siyu Ye⁷, Jinghao Lin⁷, Xiaozhong Ji⁷, Boyuan Jiang⁷, Shengzhi Wang¹, Zihan Wang⁷, Yiwen Ye⁷, Hao Wang⁷, Zimu Wang³, Wenzhe Liu⁷, Ruobing Wang⁷, Kai Cai⁷, Mingliang Xiong¹, Wen Fang¹, Mingqing Liu¹, Yifan Zhang⁴, Lei Yang⁶, Xiaobin Hu⁵, Qingwen Liu^1,†

¹Tongji University · ²Shanghai Jiao Tong University · ³UC Berkeley · ⁴University of Chinese Academy of Sciences · ⁵National University of Singapore · ⁶Nanyang Technological University · ⁷Simple Agent Lab
^*Equal contribution · ^†Corresponding author

Paper on Preprints.org AI4AI harness: start your research here Full article on simpleagentlab.com

This is the short version of Simple Agent Lab's AI4AI research write-up: a definition of the field, one takeaway per section, the figures and the closure table. The full article, in English and Chinese, is at simpleagentlab.com/ai4ai.

What is AI4AI?

AI4AI, AI for AI, means using AI systems to improve AI systems. An agent proposes a change to a model, a training pipeline, a harness or a benchmark, runs the experiment, reads the result, and repairs what failed. The survey treats this as a long-horizon loop with five stages, goal, plan, execute, feedback and repair, and asks one question of every system: how far can it carry that loop from an idea to a verified result on its own?

Recursive self-improvement, RSI, is the special case where the system being improved is the one doing the improving. Most of what is called self-improvement today is not that: a fixed harness, an outer search over scaffolds, or an agent editing its own code under an evaluator that humans wrote. The survey audits 35 systems against the five stages to say which paradigm each one belongs to.

AI4AI is a long-horizon loop pointed at an AI system

Each atomic skill is already strong on its own. The bottleneck is keeping the improvement loop running stably for a long time.

Four organizing questions: what the loop may change, why a single pass is already long-horizon, how to measure task pressure, closure and evidence, and when self-reference turns the loop inward. — Our taxonomy of the long-horizon loop.

Read this section in full ↗

What separates AutoML, AI4AI, and RSI: which stages the system runs itself

Ask which of the five stages the system runs on its own. Today, no system picks its goal.

Paradigm	Target	Self-reference	Typical closure
AutoML / NAS	architectures and hyperparameters	no	humans fix the goal and the search space; the system closes execute only
AI4AI	any part of an AI system	not required	the system closes plan and execute; feedback is mix under a human-written evaluator
Recursive self-improvement	the improvement mechanism itself	required	all four execution stages close; the goal is still given by humans

Read the recursion ladder in full ↗

Two routes to a longer reliable horizon: model and harness

On AI4AI's long-horizon tasks, the reliable horizon is not set by the model alone: among comparable models, changing the harness can matter more than changing the model.

Model-side interventions mapped onto the plan, execute, feedback, and repair stages of the loop. — Model-side interventions by stage: plan yields supervision, execute grounds in tools and state, feedback becomes step-level credit, repair rewrites the training recipe.

Read both routes in full ↗

Why composition is less reliable than its parts

Strong parts, unreliable whole. Two mechanisms open the gap, and the failures show up in three places.

Read where the failures show up ↗

Conclusion

What AI4AI lacks is not capability. It is the authority to decide.

The eve ends the first time a system, with no human-specified goal or evaluation, has successors that reliably strengthen it. We hope this taxonomy and this audit help that moment arrive a little sooner.

If reading leaves you wanting to build: RSIHub is our open-source framework for self-evolving agents, and a place to start your own AI4AI research. Issues and pull requests are welcome.

That is the short version. The full article walks through each section, with the paper's abstract, the recursion ladder, the model and harness routes, the three places the composition gap shows up, the BibTeX entry, and a Chinese translation: read it on simpleagentlab.com.

How to cite this AI4AI survey

The survey is published on Preprints.org as doi:10.20944/preprints202608.2108.v1. The BibTeX entry is on the full article page.

← Back to Simple Agent Lab