Survey · Preprints.org · 2026

A Survey of
AI for AI

From Long-Horizon Agents to Recursive Self-Improvement
Definitions, Reliable Horizons, and Open Problems

Read the survey Explore the repository

Kai Wu · Hao Lyu · Zhen Luo · +20 co-authors

TJU · SJTU · UC Berkeley · UCAS · NUS · NTU · Simple Agent Lab

PDF (original layout) ↓ RSIHub harness ↗ Blog ↗ DOI ↗ Cite ↗

The AI4AI landscape: benchmarks and environments, the long-horizon agent, and models and harnesses covered by this survey
Figure 1 · The AI4AI landscape — what this survey covers
223public papers
111benchmarks
98harness studies
26model studies
LONG-HORIZON AGENTSRECURSIVE SELF-IMPROVEMENTAGENT BENCHMARKSHARNESS DESIGNMODEL DESIGNAUTOMATED AI RESEARCH

01 · THE RESEARCH ATLAS

See the field.
Connect the systems.

One visual map for the benchmarks that measure agents, the harnesses that guide them, and the model interventions that make them improve.

FIGURE 02WHAT IS AI4AI?
A taxonomy of AI4AI: what is being improved, why long-horizon, how to measure, and when it recurses
What is AI4AI?the taxonomy · explore ↗
SYSTEM VIEWHARNESS
Evidence chain for reliable AI agent harness interventions
Harness design98 papers ↗
SYSTEM VIEWMODEL
Model-side interventions across planning, execution, feedback, and repair
Model design26 papers ↗
FIGURE 03FULL TAXONOMY

Every paper,
placed on the map.

The survey's complete taxonomy — from “What is AI4AI?” through evaluation, model design, and harness design, to the open problems — with each of the 223 catalogued papers attached to the section that cites it.

Browse the catalog ↗
Full survey taxonomy mind map from what is AI4AI through evaluation, model design, harness design, and open problems

02 · LIVE COLLECTIONS

Curated for signal,
built for discovery.

01
⌁

Weekly intelligence

The month's highest-signal papers, releases, and research news—source-checked and reranked every Monday.

Open the signal feed ↗ 02
↗

Live rankings

Citation velocity, yearly leaders, and code popularity refreshed from scholarly APIs and GitHub.

See live rankings ↗ 03
◇

Evidence audit

Traceable evidence for stage ownership, generalization, robustness, horizon, and transfer claims.

Inspect the evidence ↗

03 · REPRODUCIBLE BY DESIGN

Every claim has
a trail.

Discover, verify, classify, audit, and refresh. The catalog combines human research judgment with reproducible data and weekly automation.

Read the methodology →
Search, screening, extraction, coding, and verification workflow
PAPER NOW LIVEPREPRINT · 2026

THE COMPANION SURVEY

AI4AI Survey

From Long-Horizon Agents to Recursive Self-Improvement—Definitions, Reliable Horizons, and Open Problems

A unified map of the systems, evidence, and design choices shaping AI that can help build better AI.

Read on Preprints.org↗ Download PDF↓ Cite this work→

THE MAP KEEPS MOVING

Follow the frontier.

Star the repository for the weekly update, contribute a missing paper, or explore the structured dataset.

Star on GitHub★ Explore the data→