Literature Survey
Recursive Self-Improvement: A Literature Survey
An annotated bibliography for anyone doing actual research in this area โ the foundational theory, four recent formal surveys, thirteen named techniques in chronological order, and the narrower "AI automating AI research" subfield, each with a real summary, a link to the paper, and a linked video discussion where one could be found and verified. This is written as a working researcher's reference list, not an explainer for a general audience.
A note on video links before you read further: YouTube itself was inaccessible to directly verify page titles and channel names while researching this piece, so video links below were located via search and could not be opened to confirm the exact on-page title matches word-for-word. Where no real, specific video discussing a paper could be found at all, this article says so explicitly rather than fabricating one โ several papers below (AutoML-Zero, FunSearch, Gรถdel Agent, I.J. Good 1965) have no video link for exactly this reason.
Opening & Method
Foundational Theory
The Four Recent Surveys
Named Techniques, Chronologically
AI Automating AI Research
What the Literature Agrees and Disagrees On
Confidence Notes & Gaps
Full Reference List
Scope and Method
This survey covers three tiers of the recursive self-improvement literature: the pre-2020 theoretical foundations that named and formalized the concept; the technical systems built since 2020 that implement some version of self-improvement empirically; and the narrower, currently very active subfield of AI systems specifically automating AI research work. Papers were selected for direct relevance to self-referential or self-modifying AI systems, not merely for using the word "self-improving" in passing. Confidence levels are stated explicitly per entry โ several of the most recent 2026-dated survey papers below postdate this research and could only be characterized from search-indexed summaries, not full-text reading, and that limitation is flagged rather than hidden.
RSI Advancement Timeline
Every entry below appears with full discussion in the sections that follow. Two panels because the field's real activity is heavily back-loaded: four sparse decades of theory, then a dense two-year run of empirical systems.
Foundational theory
Named technique
AI automating AI research
Foundational theory ยท 1965โ2013
1965
I.J. Good โ "Speculations Concerning the First Ultraintelligent Machine"
Names the intelligence explosion: a machine designing better machines than itself.
2002 / 2004
Schmidhuber โ Optimal Ordered Problem Solver (OOPS)
Bias-optimal incremental universal search; the empirical precursor to the Gรถdel Machine.
2003 / 2006
Schmidhuber โ Gรถdel Machine
Self-rewriting agent that only changes its own code once it can formally prove the rewrite helps.
2013
Yudkowsky โ Intelligence Explosion Microeconomics
Formalizes RSI as a returns-on-cognitive-reinvestment production function.
Empirical systems & AI automating AI research ยท 2020โ2025
2020
AutoML-Zero
Evolves ML algorithms from near-scratch using only basic math operations.
2023 (May)
Voyager
Open-ended Minecraft agent that writes and accumulates its own skill library.
2023 (Sep)
Promptbreeder
Self-referential evolution of the prompts that do the evolving, not just the task prompts.
2023 (Oct)
STOP & Eureka
STOP: a language model recursively improving its own scaffolding code. Eureka: LLM-authored reward functions for robot skill learning.
2023 (Dec)
FunSearch
LLM-guided evolutionary search finds a new, published-worthy result in extremal combinatorics.
2024 (Aug)
The AI Scientist
End-to-end automated pipeline: idea, experiment, write-up, for real ML papers.
2024 (Oct)
Gรถdel Agent
A practical, code-level attempt to give an LLM agent Gรถdel-Machine-style self-rewriting.
2024 (Oct)
MLE-bench & AutoML-Agent
Benchmarking and agentizing the full applied-ML-engineering pipeline.
2024 (Nov)
RE-Bench
METR's benchmark comparing frontier agents against human experts on real AI R&D tasks.
2025 (Feb)
Evaluating Sakana's AI Scientist
Independent scrutiny of the 2024 AI Scientist's claimed autonomous results.
2025 (Apr)
SICA
A self-improving coding agent that edits its own codebase across iterations.
2025 (Apr)
The AI Scientist-v2
A revised pipeline after the v1 scrutiny above, with tightened claims.
2025 (May)
Darwin Gรถdel Machine
Empirically validates code self-modification against a benchmark instead of formal proof.
2025 (Jun)
AlphaEvolve
DeepMind's LLM-driven evolutionary coding agent, applied to its own production infrastructure.
2025 (Oct)
Huxley-Gรถdel Machine
The most recent named technique surveyed here; least independently confirmed entry.
Foundational Theory
Speculations Concerning the First Ultraintelligent Machine
The origin of the "intelligence explosion" concept: an ultraintelligent machine could design even better machines than itself, triggering a runaway explosion of intelligence that leaves human intellect far behind โ "the last invention man need ever make," provided the machine remains controllable enough to tell us how to keep it under control. Every later formalization of RSI, from Schmidhuber's Gรถdel Machine to the 2026 survey papers below, traces back to this single paragraph-length speculation.
๐ Paper (DOI)
๐บ No verifiable dedicated video found
Gรถdel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements
Proposes a fully self-referential agent whose initial code includes an axiomatized description of itself, its environment, and its utility function, plus a proof searcher. The machine rewrites any part of its own code the instant it can formally prove the rewrite provably increases expected utility โ giving mathematically optimal, not just locally optimal, self-improvement. The catch, which the paper is honest about: the proof search is generally intractable, making this a foundational theoretical construct rather than a deployable system.
Optimal Ordered Problem Solver (OOPS)
An early, concrete precursor to the full Gรถdel Machine: introduces bias-optimal incremental universal search, where the system reuses and organizes solutions to earlier tasks to systematically accelerate solving new ones. This is the empirical half of Schmidhuber's self-referential-search program, published before the fully formalized Gรถdel Machine paper.
๐ Springer (Machine Learning journal)
๐บ No verifiable dedicated video found
Intelligence Explosion Microeconomics
Not peer-reviewed, but the canonical foundational document from the Machine Intelligence Research Institute โ roughly 40,000 words attempting to formalize "returns on cognitive reinvestment": under what conditions would AI self-improvement accelerate versus plateau. Frames RSI as an economic production-function problem (optimization power in, intelligence out, which becomes more optimization power), a framing that directly anticipates the "recursive reproduction number" approach in the 2026 "Recursive Criticality" paper covered below.
๐ MIRI Technical Report (PDF)
๐บ No verifiable dedicated video found
The Four Recent Formal Surveys
All four of the following postdate this research's ability to read full text directly and were characterized from search-indexed abstracts and summaries only โ treat the summaries below as a starting point for your own reading, not a substitute for it.
Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops
Systematically surveys roughly 1,250 arXiv papers (2024โ2026) along two axes: what is being improved (deployment-time behavior, the trained policy, the evaluator, or the research process itself), and degree of loop closure (from human-in-the-loop refinement to fully closed autonomous loops). Also surveys the evaluator-design space โ LLM judges, process reward models, verifiers, rubrics, meta-evaluation โ as the actual bottleneck for safe closed-loop RSI, a framing worth reading against Part 2 of the objective-specification literature.
๐ arXiv:2607.07663
๐บ No verifiable dedicated video found
The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement
Surveys 491 papers against a 5-level maturity ladder (L1 through L5) for self-improvement claims. The central, striking finding: 75.4% of surveyed work sits at L1โL2 (basic self-refinement or in-context iteration), while only 5.7% reaches L4 (deployment-time self-adaptation) and 5.9% reaches L5 (meta-level improvement of the improvement process itself). If accurate, this is strong evidence that genuine RSI remains rare in the literature despite how often the term gets applied loosely.
๐ arXiv:2609.11873
๐บ No verifiable dedicated video found
Self-Improvements in Modern Agentic Systems: A Survey
Broader in scope than pure RSI โ surveys self-referential neural architectures and self-improvement mechanisms across modern agentic AI systems generally, including self-refinement, self-play, and memory-based adaptation alongside stricter recursive self-improvement. Useful as a map of how RSI proper relates to the wider "agentic self-improvement" literature it's often conflated with.
๐ arXiv:2607.13104
๐บ No verifiable dedicated video found
Recursive Criticality of AI Self-Improvement
A more theoretical and quantitative entry than the other three surveys โ models the rate of AI capability growth as a function of baseline research productivity, the strength of recursive feedback, and the increasing marginal difficulty of further research progress, deriving an epidemiology-style "recursive reproduction number" for when self-improvement becomes explosive versus self-limiting. Directly extends the economic framing Yudkowsky introduced in 2013 into a more formal model.
๐ arXiv:2609.00137
๐บ No verifiable dedicated video found
A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence
The most-cited general taxonomy paper in this space as of late 2025: the first systematic taxonomy of "self-evolving agents" along four dimensions (what to evolve, when to evolve, how to evolve, where to evolve), and the first to clearly distinguish intra-test-time evolution (real-time adaptation within a single task) from inter-test-time evolution (updates that persist across tasks and episodes) โ a distinction worth applying to every technique paper in the next section.
๐ arXiv:2507.21046
๐บ No verifiable dedicated video found
Named Techniques, Chronologically
AutoML-Zero: Evolving Machine Learning Algorithms From Scratch
Evolves entire ML algorithms โ not just architectures or hyperparameters โ from primitive mathematical operations, using regularized evolution over a generic instruction-based search space. Notably rediscovers real concepts like backpropagation, gradient normalization, bilinear interactions, and dropout-like regularization without any of them being pre-specified in the search space, making it one of the earliest credible demonstrations that a search process can reinvent known ML building blocks from near-first-principles.
๐ arXiv:2003.03384
๐บ No verifiable dedicated video found
Promptbreeder: Self-Referential Self-Improvement via Prompt Evolution
Evolves a population of task-prompts using LLM-generated mutation-prompts that are themselves evolved self-referentially โ the mutation strategy improves alongside the prompts it's mutating. Outperforms Chain-of-Thought and Plan-and-Solve prompting on reasoning benchmarks without any human prompt-engineering after the initial seed.
STOP: Self-Taught Optimizer โ Recursively Self-Improving Code Generation
A seed "improver" program โ one that calls an LLM multiple times to improve a candidate program against a utility function โ is applied to improve itself. The resulting improved improvers discover real optimization strategies (beam search, genetic algorithms, simulated annealing) without being told to. The authors are explicit about the limitation this article treats as important throughout: since the underlying LLM's weights never change, this is scaffolding-level self-improvement, not weight-level RSI.
๐ arXiv:2310.02304
๐บ Candidate videos found but channel/title unverifiable โ see confidence notes
Voyager: An Open-Ended Embodied Agent with Large Language Models
A Minecraft agent with an automatic curriculum, an ever-growing executable skill library, and iterative GPT-4 prompting that incorporates environment feedback and self-verification. Skills compound over time without any weight updates, sidestepping catastrophic forgetting entirely by storing improvement in an external, growing code library rather than in the model itself โ a design pattern that reappears in SICA and the Darwin Gรถdel Machine below.
๐ arXiv:2305.16291
๐บ Candidate videos found but channel/title unverifiable โ see confidence notes
Eureka: Human-Level Reward Design via Coding Large Language Models
Uses GPT-4 as an evolutionary reward-function designer for robotics RL: it generates reward code, the code gets evaluated via GPU-parallelized training in Isaac Gym, and the LLM iterates using the resulting training statistics as textual feedback. Outperformed human-expert-written reward functions on 83% of a 29-task robotics benchmark โ a rare case where an LLM-in-the-loop system beat domain-expert humans on a task those humans specialize in.
FunSearch: Mathematical Discoveries from Program Search with Large Language Models
Pairs a frozen, unmodified pretrained LLM (which proposes candidate programs) with an automated evaluator in an evolutionary loop over a growing program database. Found genuinely new constructions for the cap set problem and improved bin-packing heuristics โ the first case of an LLM-driven system producing new results on an open mathematics problem, published in Nature rather than merely claimed in a preprint.
๐ Nature 625, 468โ475
๐บ No verifiable dedicated video found
Gรถdel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement
Explicitly Gรถdel-Machine-inspired but empirical rather than proof-based: an LLM agent that can inspect and rewrite its own logic and code โ including the code that does the inspecting and rewriting โ guided only by a high-level objective stated in the prompt, without a fixed meta-optimization algorithm imposed from outside. Reports roughly 11% improvement on complex reasoning tasks over manually crafted meta-learning baselines.
๐ arXiv:2410.04444
๐บ No verifiable dedicated video found
AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery
A Gemini-powered evolutionary coding agent pairing LLM-generated code mutations with automated evaluators. Found a 4ร4 complex matrix multiplication algorithm using 48 scalar multiplications โ the first improvement over Strassen's 1969 algorithm in this specific setting in over five decades โ and matched or exceeded best-known solutions on a large fraction of a battery of open math problems. Notably, it was also applied recursively to improve Google's own data-center scheduling, chip design, and LLM training pipelines, including the training of the very Gemini models AlphaEvolve itself runs on โ a genuine, disclosed feedback loop rather than a hypothetical one.
Darwin Gรถdel Machine: Open-Ended Evolution of Self-Improving Agents
Explicitly built on Schmidhuber's Gรถdel Machine but replaces the intractable formal proof-of-improvement requirement with empirical validation: maintains a growing archive of agent versions, samples from that archive, and lets the sampled agent self-modify its own Python codebase โ its own tools, context management, and review mechanisms. Raised SWE-bench performance from 20.0% to 50.0% and Polyglot benchmark performance from 14.2% to 30.7% purely through iterative self-modification, without any human editing the code between iterations.
๐ arXiv:2505.22954
๐ป Code (GitHub)
๐บ Candidate videos found but channel/title unverifiable โ see confidence notes
A Self-Improving Coding Agent (SICA)
A fully self-referential design: the same agent that solves coding tasks is the one that edits its own entire codebase โ its tools, sub-agents, and orchestration logic โ via a "Meta-Agent Loop" in which the best archived version of the agent acts as the improver for the next iteration. Improved a subset of SWE-bench Verified from 17% to 53% through this process alone.
๐ arXiv:2504.15228
๐บ Candidate videos found but channel/title unverifiable โ see confidence notes
Huxley-Gรถdel Machine
A follow-on to the Darwin Gรถdel Machine claiming further progress toward human-level coding-agent self-improvement. This entry's author list and precise technical contribution were not independently verified in depth during research for this piece โ included for completeness, but readers should treat it as the least-confirmed entry in this section pending direct reading.
๐ arXiv:2510.21614
๐บ No verifiable dedicated video found
AI Automating AI Research
The narrower subfield this survey was specifically asked to focus on: papers about AI systems automating the process of AI/ML research itself โ ideation, experiment design, execution, and writing up results โ rather than self-modifying code in the abstract.
The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery
An end-to-end automated research pipeline: idea generation, literature check, experiment execution, paper writing, and automated LLM-based peer review, applied to small machine learning research topics such as diffusion models and grokking. Best generation results used Claude 3.5 Sonnet, with GPT-4o performing the review step; the authors claim roughly $15 per generated paper.
Evaluating Sakana's AI Scientist for Autonomous Research
A deliberate skeptical counterweight to include alongside The AI Scientist itself: an independent evaluation specifically questioning the actual research quality and validity of AI Scientist's outputs, asking whether the system represents genuine progress toward "Artificial Research Intelligence" or overstates what it actually achieves. Worth reading paired with the original paper rather than in isolation.
๐ arXiv:2502.14297
๐บ No verifiable dedicated video found
The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search
Removes v1's reliance on human-authored code templates entirely, replacing it with a progressive agentic tree-search methodology and a dedicated experiment-manager agent that incorporates vision-language-model feedback on generated figures. Produced the first fully AI-generated paper to pass peer review at an ICLR 2025 workshop โ 1 of 3 submitted papers cleared the acceptance threshold, a specific, checkable claim rather than a vague one.
RE-Bench: Evaluating Frontier AI R&D Capabilities of Language Model Agents Against Human Experts
A benchmark of seven open-ended ML research-engineering environments (e.g., fitting scaling laws, optimizing GPU kernels), compared directly against 71 eight-hour attempts by 61 human experts โ one of the few papers in this survey with a genuine human baseline rather than only a model-to-model comparison. The widely cited finding, worth verifying the exact framing against the paper's own tables before repeating it elsewhere: the best AI agents scored roughly 4x higher than human experts when both were given a 2-hour budget, though the human/AI comparison narrows or reverses at longer time budgets.
MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering
75 Kaggle-derived ML engineering competitions with real human leaderboard baselines to compare against. The best tested setup (o1-preview with AIDE scaffolding) achieved at least Kaggle-bronze-medal-equivalent performance in 16.9% of competitions. Explicitly framed by OpenAI as part of its Preparedness Framework work on measuring AI-autonomy and AI-R&D-automation risk, not as a pure capability showcase.
AutoML-Agent: A Multi-Agent LLM Framework for Full-Pipeline AutoML
A multi-agent LLM system automating the full AutoML pipeline โ retrieval-augmented planning, parallel specialized sub-agents, and multi-stage verification โ end to end. Distinct from the self-modifying-code systems above: this automates the ML engineering workflow rather than modifying its own underlying code, making it a useful contrast case for where "automating AI research" and "recursive self-improvement" diverge as categories.
๐ arXiv:2410.02958
๐บ No verifiable dedicated video found
What the Literature Actually Agrees and Disagrees On
Point of Convergence
Nearly every technique paper in this survey that reports genuine iterative improvement โ STOP, Voyager, SICA, the Darwin Gรถdel Machine โ is explicit, in its own text, that the underlying model's weights are not being modified. Improvement lives in external, growing artifacts: a codebase, a skill library, an archive of agent versions. This is a real and recurring pattern across a decade of otherwise very different systems, not a coincidence.
Point of Disagreement
The four 2026 survey papers don't fully agree on taxonomy. "The Last AI Built by Humans" uses a 5-level maturity ladder; "Recursive Self-Improvement in AI" uses a two-axis (what/how-closed) framework; "A Survey of Self-Evolving Agents" uses a four-dimension (what/when/how/where) taxonomy. These aren't contradictory so much as differently sliced, but a researcher citing "the" taxonomy for this field should be aware there isn't yet a single agreed-upon one โ a genuinely young field still converging on shared vocabulary.
The Most Consequential Single Finding
If "The Last AI Built by Humans" survey's numbers hold up under scrutiny โ 75.4% of surveyed self-improvement papers at the two lowest maturity levels, only 5.9% reaching genuine meta-level self-improvement โ it suggests the term "self-improving AI" is being applied far more loosely in the broader literature than the handful of genuinely rigorous systems in this survey's technique section warrant. That gap between labeling and substance is worth keeping in mind when reading any new paper that claims RSI without demonstrating which level it actually reaches.
Continuing this research? A follow-up piece, 2026 in RSI Research: New Breakthroughs and Papers, covers everything published since this survey was written โ new named techniques, new surveys, the 2026 AI-automating-AI-research benchmark wave, and what frontier labs have reported about self-improvement in production. Want to work hands-on instead? A Practitioner's Primer on RSI covers what to read first, real repos to run, and how to design and publish your own experiments.
โ ๏ธ Confidence Notes and Gaps
This survey's confidence varies significantly by entry, and that variance is worth restating in one place. High confidence, multiply-sourced and cross-checked: the Gรถdel Machine, OOPS, I.J. Good 1965, Yudkowsky 2013, AutoML-Zero, Voyager, Eureka, STOP, FunSearch, AlphaEvolve, the Darwin Gรถdel Machine, SICA, Gรถdel Agent, both AI Scientist papers, RE-Bench, MLE-bench, Promptbreeder, and AutoML-Agent. Lower confidence, flagged individually above: the exact author list and technical contribution of the Huxley-Gรถdel Machine paper, the precise phrasing of RE-Bench's human/AI crossover claim at longer time budgets, and all four 2026-dated survey papers, which were characterized from search-indexed abstracts rather than full-text reading and should be independently verified before being cited as settled characterizations in further work. Video links throughout could not be opened directly to confirm exact on-page titles and channel names (YouTube itself was inaccessible during research for this piece) โ treat every video link as "a real, findable video that appears to be about this paper" rather than "independently confirmed word-for-word."
๐ Full Reference List
๐ All Papers, in Order of Appearance
- Good (1965) โ "Speculations Concerning the First Ultraintelligent Machine"
- Schmidhuber (2003/2006) โ Gรถdel Machines, arXiv:cs/0309048
- Schmidhuber (2004) โ Optimal Ordered Problem Solver
- Yudkowsky (2013) โ Intelligence Explosion Microeconomics
- "Recursive Self-Improvement in AI" survey, arXiv:2607.07663
- "The Last AI Built by Humans" survey, arXiv:2609.11873
- "Self-Improvements in Modern Agentic Systems" survey, arXiv:2607.13104
- "Recursive Criticality of AI Self-Improvement," arXiv:2609.00137
- "A Survey of Self-Evolving Agents," arXiv:2507.21046
- AutoML-Zero, arXiv:2003.03384
- Promptbreeder, arXiv:2309.16797
- STOP, arXiv:2310.02304
- Voyager, arXiv:2305.16291
- Eureka, arXiv:2310.12931
- FunSearch, Nature 625
- Gรถdel Agent, arXiv:2410.04444
- AlphaEvolve, arXiv:2506.13131
- Darwin Gรถdel Machine, arXiv:2505.22954
- SICA, arXiv:2504.15228
- Huxley-Gรถdel Machine, arXiv:2510.21614
- The AI Scientist, arXiv:2408.06292
- Evaluating Sakana's AI Scientist, arXiv:2502.14297
- The AI Scientist-v2, arXiv:2504.08066
- RE-Bench, arXiv:2411.15114
- MLE-bench, arXiv:2410.07095
- AutoML-Agent, arXiv:2410.02958