A Survey · 3DAgentWorld
When the Improver Becomes the Improvee
The Review in two paragraphs. Read the full text for the complete version.
AI for AI concerns processes that use AI to improve AI systems and their development methods. This Review examines how AI contributes to these improvement processes, including model learning and harness engineering. We focus on cases where the improver itself changes. We compare updates to experience generators, evaluation mechanisms and programs that produce improvements. A system may combine these paths. Their effects depend on the feedback used, the experience kept for later rounds and the external checks that remain fixed.
We examine their benefits, failure modes and costs. We distinguish changes to an improver from gains in its ability to improve AI in later rounds. We then ask when improvement becomes recursive, how gains scale with resources, how long progress can continue and whether improvers generalize across AI systems.
Continue reading →Four moves that turn a scattered literature into one testable question.
Methods grouped by the work the improver supplies: training experience, evaluation feedback, or changes to prompts, workflows and code. One system may take several roles, and the improver and improvee need not share a form.
The paths by which generators, evaluators and modification programs themselves change, and what each path keeps for later rounds: selected traces, stored feedback, rubrics, reward code or archived agents.
Coverage narrowing under success-based selection, evaluator drift, error accumulation in memory and archives, and the transfer and cost questions that task scores alone cannot answer.
Four open questions close the Review: when improvement becomes recursive, whether gains follow a scaling law, what limits the improvement horizon, and whether one improver can generalize across improvees under comparable budgets and independent tests.
Every method in the Review supplies one or more of three kinds of work. See Figure 1 and Section 3 for the full account.
Browse all 195 papers behind the survey, including publication links, taxonomy placement, and 108 papers with available code.
Seven sections, from the anatomy of improvement to an evaluation agenda for updated improvers.
If this survey is useful to your work, please cite it as follows.
@article{202609.0591,
doi = {10.20944/preprints202609.0591.v2},
url = {https://doi.org/10.20944/preprints202609.0591.v2},
year = 2026,
month = {September},
publisher = {Preprints},
author = {Deheng Ye and Zheng Zhang and Hao Wang and Chunyan Miao},
title = {A Survey on AI for AI: When the Improver Becomes the Improvee},
journal = {Preprints}
}