AI4AILearning
Target parameters are updated.
Training procedures, architectures, optimizers, and parameter-updating post-training.
Configurations & training pipelinesArchitectures & update rulesParameter-updating post-training
ARTIFICIAL INTELLIGENCE FOR ARTIFICIAL INTELLIGENCE
: Toward AI-discovered
Superintelligence
Not just what AI does. What AI decides.
7 institutions·† Core contribution·Corresponding authors
1Nankai University2Zhongguancun Academy3Xi'an Jiaotong University4Shanghai Jiao Tong University5Nanyang Technological University6National University of Singapore7The Hong Kong University of Science and Technology
Contacts-lsyuan25@bza.edu.cn
3 control decisions4 artifact families336 references
01 THE BIG PICTURE
A new lens on AI-built AI.
THE AI4AI LANDSCAPE
The progressive transfer of control over the AI4AI discovery loop.
In the paper ↗THE CENTRAL QUESTION
As AI takes on more research,
which decisions does it
actually control?
IntroductionExecuting an entire research workflow does not necessarily mean controlling it. Humans may still define the goals, the boundaries of allowable change, and the criteria for success.
This survey follows the transfer of three discovery decisions: search direction, search space, and evaluation signal. It connects that control-centered view to four families of AI artifacts.
Read the full abstractABSTRACT
Page 1Artificial intelligence (AI) is reshaping how AI itself is developed. From designing learning algorithms to automating research workflows and evaluation, this expanding role has fueled the rapid growth of AI for AI (AI4AI). Yet a fundamental question remains: as AI takes on more research tasks, which decisions does it actually control? Even systems that execute entire research workflows may still rely on humans to define research goals, allowable modifications, and criteria for success. Existing taxonomies, typically organized by technique or target artifact, offer limited insight into these differences in research responsibility. This survey presents a unified, control-centered perspective on AI4AI, viewing its evolution as a progressive transfer of control over three key decisions in the discovery loop: search direction, search space, and evaluation signal. We introduce an L0–L3 control-source framework, complemented by full responsibility profiles, to characterize different forms of human–AI research responsibility without conflating control with overall capability or end-to-end autonomy. This perspective unifies research across AI4AI Learning, Inference, Data, and Infrastructure, while motivating an evaluation framework that goes beyond downstream performance to assess decision quality, discovery efficiency, transfer, evaluator independence, and robustness under optimization. Looking further ahead, AI4AI may enable recursive self-improvement, where advances in AI enhance the process of making subsequent discoveries. If sustained and independently validated, such improvements could offer a pathway toward superintelligence. By connecting research autonomy with the conditions for reliable and cumulative progress, this survey provides a unified foundation for understanding the evolution and long-term potential of AI4AI.
KEYWORDSAI for AI · Automated AI Research · Research Autonomy
02 THE CONTROL-SOURCE FRAMEWORK
Three decisions.
Three transfers of control.
L1–L3 describe the source of research control,
not a ranking of intelligence or capability.
THE CONTROL FRAMEWORK
The three control variables and the evidence-driven discovery loop. Direction, search space, and evaluation are attributed separately.
In the paper ↗ L1 · AI-PROVIDED SEARCH DIRECTION
Which direction should we pursue next?
AI uses previous candidates, scores, critiques, failures, and experimental records to select or revise the next improvement direction. The search space and evaluation signal remain human-authored or externally fixed.
Read Section 4RESPONSIBILITY PROFILE · D, S, E⟨A, H, H⟩
A AI-provided H Human / externally fixed * Either source
What remains anchored?Search space + evaluation signal
L1 · SEARCH DIRECTION
History-conditioned direction selection at L1. AI decides what to try next within an externally fixed search space and evaluation criterion.
In the paper ↗01Performance histories02Evolutionary histories03Reflective histories04Experimental histories
L1 · RESEARCH TAXONOMY
Scroll sideways to read the details, or open the full figure.
Taxonomy of L1 search-direction discovery from accumulated history: performance, evolutionary, reflective, and experimental histories.
In the paper ↗Explore representative work
L2 · AI-PROVIDED SEARCH SPACE
What kinds of solutions can we consider?
AI constructs or substantively revises the representations, primitives, composition and edit operators, or validity constraints that define the active search space. Direction may come from either source; evaluation remains external.
Read Section 5RESPONSIBILITY PROFILE · D, S, E⟨*, A, H⟩
A AI-provided H Human / externally fixed * Either source
What remains anchored?Evaluation signal
L2 · SEARCH SPACE
AI construction of the active search space at L2. AI can revise representations, primitives, operators, and admissibility constraints; evaluation remains externally fixed.
In the paper ↗01Candidate representations02Available primitives03Composition & edit operators04Admissibility constraints
L2 · RESEARCH TAXONOMY
Scroll sideways to read the details, or open the full figure.
Component-based taxonomy of L2 AI-provided search spaces. Attribution applies to the component AI constructs or revises.
In the paper ↗Explore representative work
L3 · AI-PROVIDED EVALUATION SIGNAL
How do we decide whether an idea is better?
AI constructs or substantively revises the operative tests, judgments, aggregation rules, or decision thresholds that govern selection, training, or acceptance. Direction and search space may remain human-provided.
Read Section 6RESPONSIBILITY PROFILE · D, S, E⟨*, *, A⟩
A AI-provided H Human / externally fixed * Either source
What remains anchored?Human anchors can remain
01Executable evaluation artifacts02Learned preference evaluation & model judges03Generated objectives & self-evaluation04Adaptive evaluation & oversight
L3 · EVALUATION SIGNAL
AI-provided evaluation signals in the discovery loop. Independent validation audits gains separately from the operative signal used for optimization.
In the paper ↗L3 · RESEARCH TAXONOMY
Scroll sideways to read the details, or open the full figure.
Taxonomy of AI-provided evaluation artifacts. These are L3 candidates; an L3 assignment requires the signal to become operative in the discovery loop.
In the paper ↗Explore representative work
L0The human-controlled reference⟨H, H, H⟩
All three decisions are human-authored or externally fixed.
TRY THE FRAMEWORK
Who supplies each decision?
Switch the source to explore all eight responsibility profiles.
L1⟨A, H, H⟩
AI-provided search directionAI controls direction; search space and evaluation remain human-provided or externally fixed.
Classification precedence: E > S > D. Only ⟨A, A, A⟩ places all three decisions under AI control. Section 3.6
03 AN INDEPENDENT AXIS
Across four artifact families.
The control profile tells us who decides. The artifact family tells us what improves.
Keep both axes visible when comparing AI4AI systems.
APPLICATION OVERVIEW
Overview of AI4AI applications organized by the target of improvement: Learning, Inference, Data, and Infrastructure.
In the paper ↗APPLICATION TAXONOMY
Scroll sideways to read the details, or open the full figure.
AI4AI applications grouped by the target of improvement: Learning, Inference, Data, and Infrastructure.
In the paper ↗AI4AILearning
AI4AIInference
Target parameters stay frozen.
Prompts, workflows, agent scaffolds, and test-time reasoning or serving policies.
Prompt & workflow optimizationParameter-free agent & tool designTest-time reasoning & serving
AI4AIData
Data itself is the discovery artifact.
Selected examples, mixtures, synthetic corpora, transformations, and curricula.
Selection & curationAugmentation, synthesis & curricula
AI4AIInfrastructure
The software and hardware stack.
Kernels, compiler schedules, execution plans, accelerator designs, and EDA flows.
Kernels & tensor programsExecution plans & hardware co-designLogic, RTL & physical design
The target artifact determines the family—not whether the search controller itself is trained. Section 7
04 THE RESEARCH LIBRARY
Explore the literature.
Every work the survey discusses in the L1–L3 chapters and in Applications & Benchmarks.
Click a title to open the paper; the card shows where the survey discusses it.
297Works in the level and application chapters
146Methods in the survey’s tables and taxonomies
54Benchmarks and evaluation suites
336Bibliographic references
CONTROL SOURCE
APPLICATION FAMILY
| PAPER · CLICK TO OPEN | YEAR | LEVEL · WHERE | FAMILY · TYPE | Where the survey discusses it |
|---|
No matching references.
Try a broader term or remove a filter. Works the survey does not place at a level are not given inferred levels.
Read the badges with care. “Table” levels come from Tables 3–9; “Figure 4” marks a leaf of the L1 taxonomy; “Figure 6” is component-specific L2 attribution; “Figure 8” lists L3 candidates that need closed-loop evidence; a section badge such as “§4.2” means the work is discussed in that chapter without a table or taxonomy level. Benchmarks carry no level. Titles and years follow the manuscript; article links come from the manuscript or from a title-matched arXiv or DOI record, and titles without one open a search. How attribution works
05 WHAT COMES NEXT
More control.
Harder questions.
Reliable AI4AI needs evidence about the discovery process—not just a better final score.
01When is research control worth delegating?
AI-provided direction, search space, or evaluation does not necessarily improve research quality. The value of delegation may depend on how quickly errors become observable, whether they are reversible, how costly independent verification is, and how much human effort the decision would otherwise require. Should research autonomy be a fixed configuration or a revisable allocation of authority that changes with evidence and risk?
FUTURE WORK · CAUSAL EVALUATION AND ADAPTIVE ALLOCATION OF CONTROL
Section 8.1 · page 49Expose D, S, and E as inspectable, versioned components that can be replaced independently or jointly, record human effort and rescue interventions, and test delegation policies that allocate authority by uncertainty, consequence, and verification cost.
02Can AI expand the boundaries of practical discoverability?
A sufficiently expressive language may already represent a desirable algorithm without making it discoverable within a realistic budget. The central question for search-space construction is whether AI can make previously inaccessible, high-quality solutions practically reachable, and whether it can improve how research itself is organized.
FUTURE WORK · OPEN-ENDED SEARCH-SPACE CONSTRUCTION AND META-DISCOVERY
Section 8.1 · page 49Ask systems to diagnose why an existing representation obstructs progress and propose a space-defining artifact; count the cost of constructing the space and test it on unseen tasks with several compatible search procedures.
03What establishes progress when evaluation itself evolves?
When a candidate system and its evaluator change together, rising internal scores no longer isolate capability gains from changes in measurement. A revised evaluator may catch overlooked failures, or it may reward easier tasks, favor the generator’s outputs, or relax acceptance conditions. Reliable assessment needs evidence that is not reducible to the loop’s own judgments.
FUTURE WORK · INDEPENDENTLY GROUNDED EVALUATOR CO-EVOLUTION
Section 8.1 · page 49Pair an adaptive internal evaluator with a validation path the improving system cannot rewrite during the run; vary search intensity and evaluator flexibility while measuring false acceptance, false rejection, and divergence from external outcomes.
04When do local improvements become cumulative research knowledge?
An artifact can be useful without revealing why it works or how it should inform later designs. The stronger question is whether AI4AI can turn experimental outcomes into knowledge that supports new predictions, such as executable abstractions, supported design rules, or validated negative results, and whether that contribution can be separated from more compute or longer histories.
FUTURE WORK · KNOWLEDGE-PRODUCING RESEARCH MEMORY
Section 8.1 · page 50Keep structured records that link hypotheses, implementation versions, conditions, failures, and replications; evaluate extracted knowledge by whether an independent system can implement an inferred design rule and confirm its predictions on unseen tasks.
05When does recursive self-improvement become sustained and reinforcing?
Successive improvements to an artifact do not show that the mechanism producing them has become more capable. A recursive claim needs evidence that an earlier improvement strengthens the production of later ones; sustained and accelerating improvement are distinct claims. Can gains in inference, learning, data, and infrastructure reinforce one another after integration and verification costs?
FUTURE WORK · LONGITUDINAL AND CROSS-FAMILY RECURSIVE-IMPROVEMENT EXPERIMENTS
Section 8.1 · page 50Test each generation on common targets and held-out research tasks under matched budgets, against a frozen initial improver and a history-only control; use selective rollback to check whether later gains causally depend on earlier discoveries.
06 CITE THE SURVEY
Build on this perspective.
Cite the preprint published on Preprints.org.
Version 1 · 29 October 2026.
BIBTEX · PREPRINT
@article{liu_ai4ai_survey,
title = {{AI4AI}: Toward {AI}-Discovered Superintelligence},
author = {Liu, Siyuan and Wang, Yifei and Hu, Zheng and Hai, Tong and
Ye, Ruiqing and Yang, Jinghan and Shi, Jiamin and Yang, Zhiqin and
Meng, Mingyuan and Liu, Chang and Li, Tao and Yang, Jianfei and
Zhang, He and Zheng, Nanning and Dou, Hongkun and Fu, Jingwen},
journal = {Preprints},
year = {2026},
month = oct,
doi = {10.20944/preprints202610.2999.v1},
url = {https://www.preprints.org/manuscript/202610.2999/v1},
note = {Preprint. Version 1. Posted 29 October 2026}
}Preprints.org · Version 1 · 29 Oct 2026.