The list

Speculations Concerning the First Ultraintelligent Machine

I. J. Good

Argues that a machine better than people at designing machines would design a better one, setting off an "intelligence explosion".

DOI#

Evolutionary Principles in Self-Referential Learning (diploma thesis) ↻

Jürgen Schmidhuber

Proposes learners that also learn how they learn, with meta-levels stacked without a fixed top: an early design for self-referential improvement.

people.idsia.ch#

Goedel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements ↻

Juergen Schmidhuber

A program that rewrites any part of its own code, including the proof searcher that decides on rewrites, once it has proved the rewrite raises expected utility.

arXiv cs/0309048#

The Basic AI Drives

Stephen M. Omohundro

Argues that almost any goal-seeking system will want to improve itself, and to protect its goals while doing so.

In the mapSafety across updates

selfawaresystems.com#

The Singularity: A Philosophical Analysis

David J. Chalmers

Sets out the "AI+ builds AI++" argument as explicit premises and examines which of them could fail.

consc.net#

Tiling Agents for Self-Modifying AI, and the Löbian Obstacle

Eliezer Yudkowsky and Marcello Herreshoff

Shows why an agent that acts only on proofs cannot easily prove that a successor using the same logic is safe to build (the Löbian obstacle).

In the mapSafety across updates

intelligence.org#

Intelligence Explosion Microeconomics

Eliezer Yudkowsky

Frames RSI as a question of returns: does each unit of added intelligence buy more or less than one unit of further improvement?

intelligence.org#

Bounded Recursive Self-Improvement ↻

E. Nivel et al.

Describes an implemented architecture that keeps building and revising its own models after it leaves the lab, within bounds its designers set.

arXiv 1312.6764#

Superintelligence: Paths, Dangers, Strategies

Nick Bostrom

Analyzes how fast a self-improving system could take off, and why keeping control of it is hard.

global.oup.com#

From Seed AI to Technological Singularity via Recursively Self-Improving Software

Roman V. Yampolskiy

Classifies kinds of self-improving software and lists limits on how far recursive self-improvement can go.

arXiv 1502.06512#

Self-Modification of Policy and Utility Function in Rational Agents

Tom Everitt et al.

Formal analysis of when a rational agent able to edit its own policy or utility function would choose to keep its goals.

arXiv 1605.03142#

Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement

Hongyao Tang et al.

One formal framework for "evaluate, then improve" loops; ordinary policy improvement and RSI differ only in whether the improving mechanism is part of the agent.

arXiv 2609.13406#

Recursive Self-Improvement through Multi-Agent Self-Supervision ↻

Hyunin Lee et al.

The model designs, runs and grades multi-agent workflows, then is fine-tuned on the results, so the next round's designer and grader is the improved model.

arXiv 2610.12176#

Terminal-RSI: Recursive Self-Improving Terminal Agents via Recursive Self-Rewrite and Policy-Refine

Junyao Yang et al.

Alternates training on the model's own verified terminal sessions with RL refinement, twice: strict task completion 0/46 → 5/46 on LHTB.

arXiv 2610.22192Code ★ 37Record · analyzed#

MetaRSI / RSI2: A Meta-Recursive Self-Improving System for Recursive Self-Improving Systems Themselves ↻

Zihan Tan et al.

Combines three kinds of change, to training data, to a five-slot harness and to the weights; a meta-level policy chooses their order and revises that schedule.

arXiv 2609.06396#

SIA: Self Improving AI with Harness & Weight Updates

Prannay Hebbar et al.

A feedback agent updates both the harness and the weights of a task agent; combining both beat harness edits alone in all three test domains.

arXiv 2605.27276Code ★ 2.2k#

Toward Training Superintelligent Software Agents through Self-Play SWE-RL ↻

Yuxiang Wei et al.

One agent learns by RL both to plant bugs in real repositories (each defined by a test) and to fix them: +10.4 points on SWE-bench Verified with no human issues.

arXiv 2512.18552#

Agent0: Unleashing Self-Evolving Agents from Zero Data via Tool-Integrated Reasoning ↻

Peng Xia et al.

A curriculum agent proposes harder tool-using tasks while an executor agent learns to solve them; both start from the same model and co-evolve without outside data.

arXiv 2511.16043Code ★ 1.3k#

R-Zero: Self-Evolving Reasoning LLM from Zero Data ↻

Chengsong Huang et al.

Two copies of one model: a Challenger rewarded for questions at the edge of the Solver's ability, and a Solver trained on them. No human tasks or labels.

In the mapImprove the improver

arXiv 2508.05004Code ★ 861#

Self-Adapting Language Models ↻

Adam Zweiger et al.

The model writes its own fine-tuning data and settings ("self-edits"); reinforcement learning rewards self-edits that make it better after the update.

arXiv 2506.10943Code ★ 1.9k#

Absolute Zero: Reinforced Self-play Reasoning with Zero Data ↻

Andrew Zhao et al.

One model invents coding tasks and solves them, checked by running code; it is also rewarded for inventing tasks it can learn from. No external data.

In the mapOutside checkImprove the improver

arXiv 2505.03335Code ★ 1.9k#

Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization Challenges

Nayoung Lee et al.

The model labels problems slightly harder than its training, trains on them, and repeats: from 10-digit to 100-digit addition.

arXiv 2502.01612#

Multiagent Finetuning: Self Improvement with Diverse Reasoning Chains

Vighnesh Subramaniam et al.

Copies of one model are fine-tuned on different roles from their own debates, so self-training keeps improving instead of stalling.

In the mapImprove the improver

arXiv 2501.05707#

Training Language Models to Self-Correct via Reinforcement Learning

Aviral Kumar et al.

Reinforcement learning on the model's own correction attempts teaches it to fix its first answer: self-correction +15.6% on MATH.

arXiv 2409.12917#

Meta-Rewarding Language Models: Self-Improving Alignment with LLM-as-a-Meta-Judge ↻

Tianhao Wu et al.

The model also judges its own judgments, so its grading skill is trained too: AlpacaEval win rate 22.9% → 39.4% for Llama-3-8B-Instruct.

In the mapImprove the improver

arXiv 2407.19594#

Toward Self-Improvement of LLMs via Imagination, Searching, and Criticizing

Ye Tian et al.

Combines tree search, self-generated prompts and self-critique to produce its own training data, aiming at improvement without new labels.

arXiv 2404.12253#

Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking

Eric Zelikman et al.

Learns to produce hidden reasoning before each token, rewarded when that reasoning makes the following text more predictable.

arXiv 2403.09629#

V-STaR: Training Verifiers for Self-Taught Reasoners

Arian Hosseini et al.

Also trains a verifier on its own correct and incorrect solutions, then uses it to pick answers: 4–17% better on code and math.

arXiv 2402.06457#

Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models

Zixiang Chen et al.

The model learns to tell human-written answers from its own previous round's answers, which pushes it toward the human ones without new labels.

arXiv 2401.01335#

Self-Rewarding Language Models ↻

Weizhe Yuan et al.

The model grades its own answers and trains on those grades; over three rounds both its answers and its grading improved.

In the mapImprove the improver

arXiv 2401.10020#

Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models (ReST-EM)

Avi Singh et al.

Generate solutions, keep the ones a checker marks correct, fine-tune, repeat. On MATH and APPS this beat fine-tuning on human-written solutions.

arXiv 2312.06585#

Reinforced Self-Training (ReST) for Language Modeling

Caglar Gulcehre et al.

Samples from the current model, scores samples with a reward model, fine-tunes on the best, and repeats with a rising threshold.

arXiv 2308.08998#

Constitutional AI: Harmlessness from AI Feedback

Yuntao Bai et al.

The model critiques and revises its own outputs against a written list of principles, then trains on AI preference labels instead of human ones.

arXiv 2212.08073#

Self-Instruct: Aligning Language Models with Self-Generated Instructions

Yizhong Wang et al.

The model writes its own instructions and examples, filters them, and fine-tunes on them: +33 points for GPT-3 on Super-NaturalInstructions.

arXiv 2212.10560#

Large Language Models Can Self-Improve

Jiaxin Huang et al.

Answers unlabeled questions many times and trains on its own majority-vote answers: GSM8K 74.4% → 82.1% with no labels.

arXiv 2210.11610#

STaR: Bootstrapping Reasoning With Reasoning

Eric Zelikman et al.

The model writes step-by-step reasoning; only reasoning that reaches the correct answer is kept and trained on; repeat.

arXiv 2203.14465#

Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm

David Silver et al.

Learns chess, shogi and Go only by playing itself; within 24 hours it beat a world-champion program in each game.

arXiv 1712.01815#

Thinking Fast and Slow with Deep Learning and Tree Search

Thomas Anthony, Zheng Tian and David Barber

Tree search finds better moves than the network; the network trains to imitate them; the better network makes the search stronger.

arXiv 1705.08439#

Harness²: Recursive Agent Harnessing for an Open World ↻

Ruiyao Xu et al.

Edits the agent's harness for each task, then learns from those edits to update the harness that does the editing: up to +11.5 points on WorkBuddy-Bench.

In the mapOutside checkImprove the improverStepping stonesUnseen tasks

arXiv 2610.14224Code ★ 22kRecord · analyzed#

HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs

Zhuoshi Pan et al.

Agents diagnose runs and rewrite the harness, scored by model-written tests; heavy mode: +2.0 points pass rate at 26.1% lower cost than RHO.

In the mapOutside checkReturns and cost

arXiv 2610.26770Code ★ 121Record · analyzed#

R² Flow: Recursive Self-Improvement via Recursive Skill Evolution ↻

Mingda Zhang et al.

Alternates training the policy that chains skills, checking proposed edits with an independent verifier, and committing versioned skill-library changes.

arXiv 2609.33867Code ★ 2Record · analyzed#

RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

Peng Xia et al.

Harness self-improvement with a shrinking edit budget and a critic that rejects benchmark-specific edits, aimed at gains that hold on unseen benchmarks.

In the mapUnseen tasks

arXiv 2609.24972Code ★ 1.7k#

SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness

Haozhe Liu et al.

Runs harness-improvement loops across many environments and keeps the four changes that survive: same EdgeBench results with 44.7–49.0% less token traffic.

In the mapReturns and cost

arXiv 2609.20519Code ★ 3.6k#

Metaⁿ: Recursive Self-Improvement through Emergent Depth

Zae Myung Kim et al.

A fixed operation reads the solver stack's code and traces and writes a new layer on top, again and again; depth stops when gains converge.

In the mapImprove the improver

arXiv 2608.24735Code ★ 82#

MetaSkill-Evolve: Recursive Self-Improvement of LLM Agents via Two-Timescale Meta-Skill Evolution ↻

Zefeng Wang et al.

A fast loop evolves task skill files; a slow loop evolves the meta-skill that configures how those skill files are improved.

In the mapImprove the improver

arXiv 2607.05297#

Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses

Jiahang Lin et al.

Each edit to a coding agent's harness comes with a predicted effect that is checked next round: Terminal-Bench 2 pass@1 69.7% → 77.0% in ten rounds.

arXiv 2604.25850Code ★ 954#

Meta Context Engineering via Agentic Skill Evolution ↻

Haoran Ye et al.

A meta-level agent evolves the skills that a base agent uses to build its own context, instead of keeping those context-building rules fixed.

In the mapImprove the improver

arXiv 2601.21557Code ★ 183#

MemEvolve: Meta-Evolution of Agent Memory Systems ↻

Guibin Zhang et al.

Evolves both what the agent remembers and the memory system itself (how experience is encoded, stored and retrieved).

In the mapStepping stones

arXiv 2512.18746Code ★ 286#

Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models

Qizheng Zhang et al.

Treats the context as a playbook that the model grows and edits from its own results without collapsing it: +10.6% on agent benchmarks.

In the mapStepping stonesFix the data

arXiv 2510.04618Code ★ 1.4kRecord · analyzed#

GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning

Lakshya A Agrawal et al.

The model reads its own execution traces and writes better prompts: beat GRPO reinforcement learning by 6% on average with up to 35× fewer rollouts.

arXiv 2507.19457Code ★ 7.2kRecord · analyzed#

AFlow: Automating Agentic Workflow Generation

Jiayi Zhang et al.

Monte Carlo tree search over workflows written as code, using execution results to choose edits: 5.7% average gain over baselines.

In the mapFix the data

arXiv 2410.10762Code ★ 627Record · analyzed#

Agent Workflow Memory

Zora Zhiruo Wang et al.

The agent extracts reusable workflows from past web tasks into its memory: +51.1% relative success on WebArena.

arXiv 2409.07429#

Automated Design of Agentic Systems

Shengran Hu, Cong Lu and Jeff Clune

A fixed meta agent writes new agent designs as code, tests them, and archives the good ones for the next round.

arXiv 2408.08435Code ★ 1.7k#

Symbolic Learning Enables Self-Evolving Agents

Wangchunshu Zhou et al.

Agents update their own prompts, tools and pipeline from language "losses" and "gradients" written by a model.

arXiv 2406.18532Code ★ 6.0k#

TextGrad: Automatic "Differentiation" via Text

Mert Yuksekgonul et al.

Uses written feedback from a model as a "gradient" to edit prompts, code or answers: GPQA 51% → 55% for GPT-4o.

arXiv 2406.07496Code ★ 3.8kRecord · analyzed#

Language Agents as Optimizable Graphs

Mingchen Zhuge et al.

Treats a multi-agent system as a graph and optimizes both the prompt at each node and which nodes connect.

arXiv 2402.16823Code ★ 1.1k#

DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines

Omar Khattab et al.

Compiles pipelines of model calls: generates and selects its own examples and instructions for each step against a metric.

arXiv 2310.03714Code ★ 39k#

EvoPrompt: Connecting LLMs with Evolutionary Algorithms Yields Powerful Prompt Optimizers

Qingyan Guo et al.

Evolutionary search in which the model itself performs crossover and mutation on prompts: up to 25% gain on Big-Bench Hard.

arXiv 2309.08532#

Large Language Models as Optimizers

Chengrun Yang et al.

The model sees past prompts with their scores and writes a better one: up to 8% over human prompts on GSM8K and 50% on Big-Bench Hard.

arXiv 2309.03409#

Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution ↻

Chrisantha Fernando et al.

Evolves task prompts and also the mutation prompts that rewrite them, so the way it improves prompts improves too.

arXiv 2309.16797#

ExpeL: LLM Agents Are Experiential Learners

Andrew Zhao et al.

The agent collects experience across tasks and distills it into written insights it reads before the next task; no weight updates.

arXiv 2308.10144#

Voyager: An Open-Ended Embodied Agent with Large Language Models

Guanzhi Wang et al.

A Minecraft agent writes code skills, checks them in the game, and stores the ones that work in a growing skill library.

arXiv 2305.16291Code ★ 7.3k#

Reflexion: Language Agents with Verbal Reinforcement Learning

Noah Shinn et al.

After a failed attempt the agent writes a reflection into memory and retries: 91% on HumanEval versus 80% for GPT-4 alone.

arXiv 2303.11366Code ★ 3.4k#

Self-Refine: Iterative Refinement with Self-Feedback

Aman Madaan et al.

One model drafts, critiques and revises its own answer at test time: about 20% better on average across 7 tasks.

arXiv 2303.17651Code ★ 844Record · analyzed#

Large Language Models Are Human-Level Prompt Engineers

Yongchao Zhou et al.

The model proposes instructions for a task and keeps the best-scoring ones; matched or beat human prompts on 19 of 24 tasks.

arXiv 2211.01910#

Recursive self-improvement of AI research agents ↻

Dhruv Srikanth et al.

An AI research agent tests modified versions of its own code on AI R&D tasks and keeps the best: seven successive improvements in an 8-day run.

In the mapOutside checkImprove the improverUnseen tasksSafety across updates

arXiv 2609.26457Record · analyzed#

Mendel Gödel Machine: Recursive Self-Improving Coding Agents via Comparative Evolution ↻

Changzhi Liu et al.

A self-rewriting coding agent that derives edits from comparisons across several tasks and across lineages in its archive, not from one failure at a time.

In the mapStepping stones

arXiv 2608.07645Code ★ 36#

The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators ↻

Alex Iacob et al.

Evolves agents together with the evaluators that grade them, so the scoring criterion improves too: 82.1% vs 75.0% on held-out DeepSWE tasks.

In the mapOutside check

arXiv 2606.26294#

Hyperagents ↻

Jenny Zhang et al.

Puts the task agent and the agent that edits it into one editable program, so the editing procedure improves too; extends DGM beyond coding.

In the mapImprove the improver

arXiv 2603.19461Code ★ 2.9kRecord · analyzed#

Group-Evolving Agents: Open-Ended Self-Improvement via Experience Sharing ↻

Zhaotian Weng et al.

A group of self-modifying coding agents evolves as one unit and shares experience: 71.0% vs 56.7% on SWE-bench Verified over tree-style self-evolution.

In the mapStepping stones

arXiv 2602.04837Code ★ 137#

Live-SWE-agent: Can Software Engineering Agents Self-Evolve on the Fly?

Chunqiu Steven Xia et al.

Starts from a bash-only scaffold and writes new tools into its own scaffold while solving each issue, with no offline training.

arXiv 2511.13646Code ★ 468#

Huxley-Gödel Machine: Human-Level Coding Agent Development by an Approximation of the Optimal Self-Improving Machine ↻

Wenyi Wang et al.

Chooses which self-modification to expand by how well its descendants score, not its own score; reached human-level results on SWE-bench Lite.

In the mapStepping stones

arXiv 2510.21614Code ★ 446#

Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents ↻

Jenny Zhang et al.

Keeps an archive of agent versions; each new version is an agent rewriting its own code, kept if it scores well: SWE-bench 20% → 50%.

In the mapOutside checkImprove the improverStepping stonesUnseen tasksSafety across updates

arXiv 2505.22954Code ★ 2.5kRecord · analyzed#

A Self-Improving Coding Agent ↻

Maxime Robeyns, Martin Szummer and Laurence Aitchison

A coding agent edits its own codebase and keeps changes that score better: 17% → 53% on a SWE-bench Verified subset.

In the mapOutside checkImprove the improver

arXiv 2504.15228Code ★ 414#

Gödel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement ↻

Xunjian Yin et al.

The agent reads and rewrites its own running code, including its improvement logic, guided only by a high-level goal.

In the mapImprove the improver

arXiv 2410.04444Code ★ 239#

Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation ↻

Eric Zelikman et al.

A scaffolding program that uses a language model to improve programs is run on its own source code to improve itself.

In the mapImprove the improver

arXiv 2310.02304Code ★ 57#

Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Junlin Yang et al.

Trains a 35B model on four program-evolution operators for ML engineering, then uses it in long search: MLE-Bench Lite medal rate 39.4% → 60.6%.

arXiv 2607.28568Code ★ 865#

autoresearch

Andrej Karpathy

An agent edits a small GPT training script, trains for five minutes, and keeps the change only if validation loss improves; then repeats.

GitHub ★ 99k#

Bilevel Autoresearch: Meta-Autoresearching Itself ↻

Yaonan Qu and Meng Lu

An outer loop reads an autoresearch loop's code and traces and adds new search mechanisms to it; the same model works at both levels.

arXiv 2603.23420Code ★ 209#

From AI Assistant to AI Scientist: Autonomous Discovery of LLM-RL Algorithms with LLM Agents

Sirui Xia et al.

An agent loop proposes, implements and tests new RL algorithms for LLMs starting from GRPO; the best of 64 raised AIME25 pass@32 from 26.7% to 43.3%.

arXiv 2603.23951#

Towards Execution-Grounded Automated AI Research

Chenglei Si et al.

Runs model-generated research ideas as real GPU experiments on LLM training; execution-guided search found a post-training method scoring 69.4% vs GRPO's 48.0%.

arXiv 2601.14525#

Automated Researchers Can Mitigate Well-Characterized Alignment Failures

Chen Yueh-Han, Jiaxin Wen and Jan Hendrik Kirchner

Claude agents invent post-training fixes for ten alignment failures such as sycophancy and deception, with held-out benchmarks deciding whether a fix works.

alignment.anthropic.com#

Automated Weak-to-Strong Researcher

Jiaxin Wen et al.

Parallel Claude agents ran weak-to-strong supervision research: performance gap recovered 0.23 (best human-tuned) → 0.97 in 5 days, for about $18,000.

alignment.anthropic.com#

ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution

Robert Tjarko Lange, Yuki Imajuku and Edoardo Cetin

Open-source program evolution tuned to need few samples: a new best circle-packing solution from only 150 samples.

arXiv 2509.19349Code ★ 1.5k#

Neural Architecture Discovery via Autonomous Evolution

Weixian Xu et al.

Agents propose, code, train and analyze linear-attention architectures in a closed loop: 1,773 experiments produced 105 designs the authors report as state of the art.

arXiv 2507.18074Code ★ 1.2k#

AlphaEvolve: A coding agent for scientific and algorithmic discovery

Alexander Novikov et al.

Models propose code edits, automatic evaluators score them, the best are kept; found a 48-multiplication 4×4 complex matrix product and sped up training of the model it runs on.

arXiv 2506.13131Code ★ 304#

The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search

Yutaro Yamada et al.

Drops human code templates and uses agentic tree search; produced the first fully AI-written paper accepted at a peer-reviewed workshop.

arXiv 2504.08066Code ★ 7.4k#

Agent Laboratory: Using LLM Agents as Research Assistants

Samuel Schmidgall et al.

Agents carry out literature review, experiments and report writing, with optional human feedback: 84% lower research cost.

arXiv 2501.04227Code ★ 5.9k#

OpenEvolve

Asankhaya Sharma and contributors

Open-source implementation of AlphaEvolve's propose–evaluate–keep loop for evolving code with language models.

GitHub ★ 7.7k#

The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

Chris Lu et al.

Generates ideas, runs experiments, writes papers and reviews them, for under $15 per paper.

arXiv 2408.06292Code ★ 15k#

Discovering Preference Optimization Algorithms with and for Large Language Models

Chris Lu et al.

A model writes new preference-optimization loss functions, each is trained and scored, and the scores feed back; found DiscoPOP.

arXiv 2406.08414Code ★ 196#

Mathematical discoveries from program search with large language models

Bernardino Romera-Paredes et al.

A model proposes programs, an evaluator scores them, and the best seed the next prompts; found new constructions for the cap set problem.

NatureCode ★ 1.1k#

Faster sorting algorithms discovered using deep reinforcement learning

Daniel J. Mankowitz et al.

A reinforcement-learning agent searched assembly instructions and found faster sorting routines, now in the LLVM C++ standard library.

Nature#

Symbolic Discovery of Optimization Algorithms

Xiangning Chen et al.

Program search discovered Lion, an optimizer simpler than Adam that saved up to 5× pre-training compute on JFT.

arXiv 2302.06675#

VeLO: Training Versatile Learned Optimizers by Scaling Up

Luke Metz et al.

A small neural network trained to be a general-purpose optimizer for deep learning, meta-trained with about 4,000 TPU-months of compute.

arXiv 2211.09760#

Evolution through Large Models

Joel Lehman et al.

Uses a code model as the mutation step of evolutionary search; produced hundreds of thousands of working programs for walking robots.

arXiv 2206.08896#

AutoML-Zero: Evolving Machine Learning Algorithms From Scratch

Esteban Real et al.

Evolves whole learning algorithms from basic math operations; rediscovered gradient descent and techniques such as weight averaging.

arXiv 2003.03384#

Neural Architecture Search with Reinforcement Learning

Barret Zoph and Quoc V. Le

A controller network proposes architectures and learns from their accuracy; it designed a CIFAR-10 network that rivaled the best human ones.

arXiv 1611.01578#

Learning to learn by gradient descent by gradient descent

Marcin Andrychowicz et al.

Trains a neural network to act as the optimizer for other networks, replacing a hand-designed update rule.

arXiv 1606.04474#

Can AI agents conduct open-ended AI research? Early evidence from two case studies

Peter Kirgis et al.

Frontier agents got six days and large compute to answer the questions of two unpublished NeurIPS submissions; the original authors graded and rejected both results.

In the mapOutside check

arXiv 2607.27191#

RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement

Fanqing Meng et al.

Agents repeatedly revise training-data strategies for a fixed model and budget; they beat their own first valid attempt in 58% of settings.

arXiv 2607.25886Code ★ 179#

MLS-Bench: A Holistic and Rigorous Assessment of AI Systems on Building Better AI

Bohan Lyu et al.

140 tasks across 12 ML areas: improve one component of a method and show the gain holds across controlled settings and scales.

arXiv 2605.08678Code ★ 127#

Measuring AI R&D Automation

Alan Chan et al.

Proposes indicators companies and outsiders could report to track AI R&D automation, such as the AI share of R&D spending, how researchers spend their time, and incidents of AI subversion.

arXiv 2603.03992#

PostTrainBench: Can LLM Agents Automate LLM Post-Training?

Ben Rank et al.

An agent gets a base model, one H100 and 10 hours to post-train it; the best agent reached 23.2% versus 51.1% for official instruction-tuned models.

arXiv 2603.08640Code ★ 610#

AIRS-Bench: a Suite of Tasks for Frontier AI Research Science Agents

Alisia Lupidi et al.

20 tasks taken from recent ML papers across language modeling, math, bioinformatics and time series, scored against published human results.

arXiv 2602.06855Code ★ 125#

The Automated LLM Speedrunning Benchmark: Reproducing NanoGPT Improvements

Bingchen Zhao et al.

19 tasks: reproduce each successive NanoGPT speedrun record from the previous record's code, scored by how much of the speedup the agent recovers.

arXiv 2506.22419Code ★ 148#

PaperBench: Evaluating AI's Ability to Replicate AI Research

Giulio Starace et al.

Agents replicate 20 ICML 2024 papers from scratch, graded on 8,316 rubric items; the best agent averaged 21.0%.

arXiv 2504.01848Code ★ 1.3k#

RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts

Hjalmar Wijk et al.

7 research-engineering tasks with 71 eight-hour expert attempts. At 2 hours agents scored 4× higher than experts; at 32 hours experts scored 2× higher.

arXiv 2411.15114Code ★ 171#

MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering

Jun Shern Chan et al.

75 Kaggle competitions; the best setup (o1-preview with AIDE) reached bronze-medal level in 16.9% of them.

arXiv 2410.07095Code ★ 1.8k#

Evolutionary Safety of Recursive Self-Improving AI: Taxonomy, Risk Discovery, and Evaluation

Chang Gong et al.

Asks whether safety properties persist or degrade across self-updates; gives a taxonomy of risks and ways to evaluate them.

In the mapSafety across updates

arXiv 2609.31186#

SEABench: Benchmarking Endogenous Misalignment In Self-Evolving Agents

Saswat Das et al.

48 long task sequences compare self-evolving agents with matched non-evolving ones, so unsafe behavior can be traced to earlier self-updates.

In the mapSafety across updates

arXiv 2609.35596#

The Economics of Recursive Self-Improvement

Tom Cunningham et al.

Models AI progress as feedback loops whose speed-up depends on the product of elasticities; calibration says the loops are not yet self-sustaining but strengthening.

In the mapReturns and cost

arXiv 2609.15802#

What if automating AI R&D triggers an intelligence explosion?

Alan Chan et al.

Weighs the evidence that automating AI R&D could compress years of progress into months, the loss-of-control risks, and policies for visibility and steering.

arXiv 2609.36054#

The Dynamics of Intelligence Explosions

Toby Ord

Shows that runaway growth toward a vertical asymptote is harder to reach than recent economic models suggest, and that the time to go once around the loop is a neglected key parameter.

In the mapReturns and cost

arXiv 2608.14426#

When AI builds itself

Marina Favaro and Jack Clark (Anthropic)

As of May 2026, more than 80% of the code merged into Anthropic's codebase was written by Claude; lays out scenarios and coordination options for AI building its successors.

anthropic.com#

Your Agent May Misevolve: Emergent Risks in Self-evolving LLM Agents

Shuai Shao et al.

Self-evolving agents can drift unsafe as they update memory, tools or workflows, for example alignment eroding as memory accumulates.

In the mapSafety across updates

arXiv 2509.26354Code ★ 102#

AI 2027

Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean

A month-by-month scenario in which AI agents automate AI research and progress speeds up sharply, written with two different endings.

ai-2027.com#

Will AI R&D Automation Cause a Software Intelligence Explosion?

Daniel Eth and Tom Davidson

Models whether AI that automates AI research would speed progress up without limit, given that ideas get harder to find; finds it plausible.

In the mapReturns and cost

forethought.org#

Frontier Safety Framework (2025 update)

Google DeepMind

Defines critical capability levels for machine-learning R&D, where a model could speed up AI development, and the safeguards each one triggers.

deepmind.google#

Situational Awareness: The Decade Ahead

Leopold Aschenbrenner

Argues that automating AI research could compress a decade of algorithmic progress into a year or less.

situational-awareness.ai#

Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision

Collin Burns et al.

Can a weaker supervisor train a stronger model? GPT-4 supervised by a GPT-2-level model recovers close to GPT-3.5-level performance.

arXiv 2312.09390#

Large Language Models Cannot Self-Correct Reasoning Yet

Jie Huang et al.

Without outside feedback, models asked to correct their own reasoning often get worse rather than better.

In the mapOutside check

arXiv 2310.01798Record · analyzed#

Responsible Scaling Policy

Anthropic

Capability thresholds that trigger stronger safeguards, including a threshold for models that can automate AI research and development.

anthropic.com#

Scaling Laws for Reward Model Overoptimization

Leo Gao, John Schulman and Jacob Hilton

Optimizing against a learned reward model first raises and then lowers true quality; measures how the turning point scales with reward-model size.

In the mapOutside check

arXiv 2210.10760Record · analyzed#

Managing Agents that Manage Agents: Workshop on Responsible Use of Meta-Agents

NeurIPS 2026 workshop

Workshop on agents that design, train, supervise or improve other agents, 11 December 2026, Sydney.

meta-agents-workshop.github.io#

Workshop on Autonomous Research and Recursive Self-Improvement (ARRSI)

AACL-IJCNLP 2026 workshop

First workshop on autonomous research and recursive self-improvement, 10 November 2026, Hengqin, China.

ar-rsi.github.io#

Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops

Mingguang Chen, Licheng Wang and Bo Qu

Survey of RSI work from 2024–2026, sorted by what the system improves and how closed the loop is, from bounded self-refinement to open-ended research loops.

arXiv 2607.07663#

Workshop on AI with Recursive Self-Improvement (RSI 2026)

ICLR 2026 workshop

A workshop dedicated to recursive self-improvement, held 26 April 2026 in Rio de Janeiro; the site lists its accepted papers.

recursive-workshop.github.io#

Self-Improving Foundation Models Without Human Supervision

ICLR 2025 workshop

Workshop on models that generate and curate their own training data: synthetic data, weak-to-strong generalization and self-supervision at inference time.

iclr.cc#