When AI Improves Itself

Awesome Self-Improving AI is an awesome list summarising papers, open-source code, and technical blogs on AI systems that improve themselves — organised along the three things a modern AI system is made of:

• 🧠 Model — the weights. Self-generated data, self-rewarding, self-play, zero-data RL, self-adapting updates.

• 🧰 Harness — everything wrapped around the weights: prompts and context, memory, skills, tools, workflows, the agent's own source code, and the evaluators that judge it.

• 🏗️ Infra — the stack the model runs on and is trained with: GPU kernels, compilers and serving configs, training pipelines, data engines, and the research process that produces the next model.

• 🎯 Self-improvement = S1 + S2 + S3. S1: the improvement signal is produced by the system itself or by an automated evaluator, not by fresh human labels. S2: the update is committed to a persistent component (weights, a harness artifact, an infra artifact) that is reused on future tasks. S3: the loop can iterate — the improved system is the one that produces the next improvement. Methods that satisfy only part of this are flagged in 📝 Strictness notes per section.

• 🚀 Each entry is annotated along three design axes — what is modified (weights · prompt/context · memory · skills · tools · workflow · harness code · evaluator · kernel/compiler · serving/training config · data · research process), improvement signal (self-consistency · self-judge/rubric · verifier/tests · environment reward · benchmark score · profiler/runtime · human-in-the-loop), and loop closure (L1 one-shot refinement · L2 bounded iteration with a fixed evaluator · L3 open-ended / recursive, where the improver itself is improved).

• ⚠️ Built by reading arXiv abstracts, project pages, and repos with LLM coding agents; cross-checked against the community lists in 🔗 Other Awesome Lists; manually reviewed but errors possible. PRs welcome.

• 📌 If you find this repository helpful for your research, please cite it via the "Cite this repository" button in the right sidebar of the GitHub page.

• 📅 Last updated: 2026-09-04

Taxonomy:

• 📚 Surveys, Foundations & Position Papers — Good, Gödel machines, the 2025–2026 surveys, the RSI definition debate
• 🧠 Model — 🧪 self-generated data & self-rewarding · ♟️ self-play & zero-data · 🔗 weights + harness co-evolution
• 🧰 Harness — ✍️ prompt & context · 💾 memory · 🧩 skills · 🕸️ workflow / agent-architecture search · 🔧 self-modifying harness code · ⚖️ evolving evaluators
• 🏗️ Infra — ⚙️ kernels · 🧮 compilers, serving & config · 🏭 training & data pipelines · 🧬 program & algorithm discovery
• 🔬 Automated AI Research — where all three loops close: systems that run the research process that makes the next system
• 📊 Benchmarks · ⚠️ Safety & Limits · 🛠️ Frameworks & Code · 📰 Blogs & Talks

Shorthand: RSI = recursive self-improvement · DGM/HGM = Darwin / Huxley Gödel Machine · RLVR = RL with verifiable rewards · MLE = machine-learning engineering · 📄 paper-only = no public code yet.

Updates

📢 click to expand

• 2026-09-04 — initial release: ~230 papers, 50+ code repos, 30 blogs/talks. Seeded from arXiv, the GitHub awesome lists on self-evolving agents / RSI / agent memory / kernel generation (credited below), Lilian Weng's harness-engineering reading list, and the Sep 2026 coverage of Anthropic's "When AI builds itself", Weco's AIDE², and Karpathy's autoresearch.

---

📚 Surveys, Foundations & Position Papers

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
Good 1965		1965	I. J. Good	PDF	Speculations Concerning the First Ultraintelligent Machine — coins the intelligence explosion: a machine that designs better machines
Gödel Machine		2003	IDSIA (Schmidhuber)	arXiv cs/0309048 · metalearning page	Gödel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements (Seminal) — rewrites any part of its own code once it can prove the rewrite helps; the theoretical north star every "Gödel" agent cites
POWERPLAY		2011	IDSIA (Schmidhuber)	arXiv 1112.5309	Training an Increasingly General Problem Solver by Continually Searching for the Simplest Still Unsolvable Problem — the open-ended self-invented-curriculum idea behind AZR / R-Zero
RSI Software		2015	Yampolskiy	arXiv 1502.06512	From Seed AI to Technological Singularity via Recursively Self-Improving Software — formal definitions of RSI and convergence limits
AutoML-Zero		2020.03	Google Brain (Real, Liang, So, Le)	arXiv 2003.03384 · code	AutoML-Zero: Evolving Machine Learning Algorithms From Scratch (ICML 2020) — evolutionary search rediscovers backprop from basic ops; the pre-LLM ancestor of AlphaEvolve-style algorithm discovery
LLM Self-Evolution Survey		2024.04	Alibaba / PKU (Tao et al.)	arXiv 2404.14387	A Survey on Self-Evolution of Large Language Models — the first survey to frame experience acquisition → refinement → updating → evaluation as one loop
Self-Evolving Agents Survey		2025.07	Princeton / Tsinghua et al. (Gao et al.)	arXiv 2507.21046	A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to ASI — what (model / memory / tools / architecture), when (intra- vs inter-test-time), how (reward / imitation / population)
Comprehensive Survey (EvoAgentX)		2025.08	Glasgow / EvoAgentX (Fang et al.)	arXiv 2508.07407	A Comprehensive Survey of Self-Evolving AI Agents: A New Paradigm Bridging Foundation Models and Lifelong Agentic Systems — single-agent / multi-agent / domain-specific optimisation taxonomy
Adaptation of Agentic AI		2025.12	UIUC (Jiang, Lin, …)	arXiv 2512.16301	Adaptation of Agentic AI: A Survey of Post-Training, Memory, and Skills — the survey that uses exactly this repo's split: weights vs memory vs skills
Externalization Review		2026.04	—	arXiv 2604.08224	Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering — argues capability is moving out of the weights into the harness
Agent System & Harness Design		2026.06	Guo, Hao et al.	arXiv 2606.20683	From Question Answering to Task Completion: A Survey on Agent System and Harness Design
Workflow Optimisation Survey		2026.03	Yue, Bhandari et al.	arXiv 2603.22386	From Static Templates to Dynamic Runtime Graphs: A Survey of Workflow Optimization for LLM Agents
Kernel Generation Survey		2026.01	FlagOS (Yu, Zang, …)	arXiv 2601.15727	Towards Automated Kernel Generation in the Era of LLMs — LLM4Kernel (SFT/RL) vs Agent4Kernel (learning · memory · profiling · multi-agent)
Measuring AI R&D Automation		2026.03	Chan, Padarath et al.	arXiv 2603.03992	Measuring AI R&D Automation — what would count as evidence that AI is doing AI research
Bounded vs Open-Ended RSI		2026.07	Chen, Wang	arXiv 2607.07663	Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops — makes loop closure the defining axis; four improvement targets (deployment behaviour, training policy, evaluator, research process)
Self-Improvements in Agentic Systems		2026.07	Ren, Chen et al.	arXiv 2607.13104	Self-Improvements in Modern Agentic Systems: A Survey — defines self-improvement as a self-induced update operator over parameters or scaffold; foundation-model vs scaffolding improvement
Co-Evolution Survey		2026.08	Zong, Liu et al.	arXiv 2608.10299	Co-Evolution in Agentic Systems: Toward Self-Directed Evolution Beyond Human Design — first survey organised around agent ↔ evaluator ↔ environment co-evolution

📝 Strictness notes (against S1 self-produced signal + S2 persistent update + S3 iterable loop)

• Good / Gödel Machine / Yampolskiy are theory: no implementation, kept as the reference definitions. The RSI definition comparison in Persdre's list shows that of the canonical sources only two even use the word "recursive".
• AutoML-Zero improves ML algorithms, not itself; listed because AlphaEvolve-style discovery descends from it.
• Surveys disagree on the top-level split: model-vs-scaffold (2607.13104), what/when/how (2507.21046), post-training/memory/skills (2512.16301). This list uses model / harness / infra because that is how the artifacts are actually deployed.

---

🧠 Model — 🧪 Self-Generated Data & Self-Rewarding

The weights are updated on data, labels, or rewards the model produced itself.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
STaR		2022.03	Stanford / Google (Zelikman, Wu, Mu, Goodman)	arXiv 2203.14465	STaR: Bootstrapping Reasoning With Reasoning (Seminal · NeurIPS 2022) — generate rationales, keep the ones that reach the right answer, fine-tune, repeat; the modern weight-level bootstrap
LMSI		2022.10	Google (Huang et al.)	arXiv 2210.11610	Large Language Models Can Self-Improve (EMNLP 2023) — self-consistency-majority answers as pseudo-labels for unlabeled questions
Constitutional AI		2022.12	Anthropic (Bai et al.)	arXiv 2212.08073	Constitutional AI: Harmlessness from AI Feedback — self-critique + revision, then RL from AI feedback; the template for replacing human labels with model judgements
Self-Instruct		2022.12	UW (Wang et al.)	arXiv 2212.10560	Self-Instruct: Aligning Language Models with Self-Generated Instructions (ACL 2023) — bootstrap an instruction dataset from the model itself
ReST		2023.08	Google DeepMind (Gulcehre et al.)	arXiv 2308.08998	Reinforced Self-Training (ReST) for Language Modeling — Grow (sample) / Improve (filter + train) loop
RLAIF vs RLHF		2023.09	Google (Lee et al.)	arXiv 2309.00267	RLAIF vs. RLHF: Scaling RL from Human Feedback with AI Feedback (ICML 2024) — AI-labelled preferences match human ones at scale
ReST-EM		2023.12	Google DeepMind (Singh et al.)	arXiv 2312.06585	Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models (TMLR) — expectation-maximisation view of STaR-style self-training; scales past human data on math/code
SPIN		2024.01	UCLA (Chen et al.)	arXiv 2401.01335	Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models (ICML 2024) — the model discriminates its own generations from human data, DPO-style
Self-Rewarding LMs		2024.01	Meta FAIR / NYU (Yuan et al.)	arXiv 2401.10020	Self-Rewarding Language Models (ICML 2024) — the model is its own reward model via LLM-as-a-judge; iterative DPO improves both policy and judge
V-STaR		2024.02	Mila / Google (Hosseini et al.)	arXiv 2402.06457	V-STaR: Training Verifiers for Self-Taught Reasoners (COLM 2024) — use the incorrect self-generated solutions to train a verifier
Quiet-STaR		2024.03	Stanford (Zelikman et al.)	arXiv 2403.09629	Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking (COLM 2024) — learn per-token internal rationales from LM likelihood alone
Meta-Rewarding		2024.07	Meta FAIR (Wu, Yuan, …)	arXiv 2407.19594	Meta-Rewarding Language Models: Self-Improving Alignment with LLM-as-a-Meta-Judge — the model also judges its own judgements, so the reward signal improves too
Self-Evolved Reward Learning		2024.11	Huang, Fan et al.	arXiv 2411.00418	Self-Evolved Reward Learning for LLMs — the reward model labels its own training data iteratively
rStar-Math		2025.01	Microsoft (Guan, Zhang, …)	arXiv 2501.04519	rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking (ICML 2025) — four rounds of MCTS-generated, code-verified data co-evolve policy and process reward model; 7B reaches o1-level MATH
Self-Improving VLM Judges		2025.12	Lin, Hu et al.	arXiv 2512.05145	Self-Improving VLM Judges Without Human Annotations — the judge bootstraps its own preference data
EvoLM		2026.05	UW (Li, Xin, … Tsvetkov)	arXiv 2605.03871	EvoLM: Self-Evolving Language Models through Co-Evolved Discriminative Rubrics — the rubric that scores generations evolves alongside the policy
Autodata		2026.06	Meta FAIR (Kulikov, Whitehouse, …)	arXiv 2606.25996	Autodata: An Agentic Data Scientist to Create High Quality Synthetic Data — an agent that designs, generates and validates the synthetic training data for the next model

📋 Click to view technical details

Resource	What is modified	Improvement signal	Loop closure	Notes
STaR	weights (SFT)	answer correctness (verifier)	L2	rationalisation on failures
LMSI	weights (SFT)	self-consistency majority	L2	no verifier needed
Constitutional AI	weights (SFT + RL)	self-critique vs a written constitution	L1→L2	human writes the constitution once
ReST / ReST-EM	weights	reward model / binary correctness	L2	grow–improve iterations
SPIN	weights (DPO)	discriminating own vs human data	L2	needs a human SFT set as the "real" class
Self-Rewarding / Meta-Rewarding	weights (DPO)	LLM-as-judge on own outputs	L3 (judge improves too)	reward hacking risk rises with iterations
rStar-Math	policy + PRM	code-execution verification + MCTS	L2	4 rounds
EvoLM	weights + rubric	co-evolved rubric	L3	evaluator is a moving target

📝 Strictness notes

• Constitutional AI and RLAIF need a human-written constitution or prompt once; after that the signal is model-produced (S1 holds, S3 only weakly).
• SPIN needs a fixed human SFT set as the positive class, so it converges rather than improving open-endedly.
• The cautionary results for this whole section live in ⚠️ Safety & Limits: Sharpening (2412.01951), Mind the Gap (2412.02674), Cannot Self-Correct Yet (2310.01798) and the model-collapse papers.

---

🧠 Model — ♟️ Self-Play, Zero-Data RL & Self-Adapting Weights

The model invents its own tasks, plays against itself, or writes its own weight updates.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
SPAG		2024.04	Tencent AI Lab (Cheng et al.)	arXiv 2404.10642	Self-playing Adversarial Language Game Enhances LLM Reasoning (NeurIPS 2024) — attacker/defender word game as a self-play RL signal
TTRL		2025.04	Tsinghua / Shanghai AI Lab (Zuo et al.)	arXiv 2504.16084	TTRL: Test-Time Reinforcement Learning (NeurIPS 2025) — majority vote over samples as the reward on unlabeled test data; RL at test time
Absolute Zero		2025.05	Tsinghua LeapLab (Zhao et al.)	arXiv 2505.03335	Absolute Zero: Reinforced Self-play Reasoning with Zero Data (NeurIPS 2025) — one model proposes code-reasoning tasks and solves them; a Python executor is the only ground truth
SEAL		2025.06	MIT (Zweiger, Pari, … Agrawal)	arXiv 2506.10943	Self-Adapting Language Models (NeurIPS 2025) — the model writes its own fine-tuning data and update directives ("self-edits"); an RL outer loop rewards edits that improve downstream performance
SPIRAL		2025.06	NUS / Sea AI Lab (Liu, Guertler, …)	arXiv 2506.24119	SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning via Multi-Agent Multi-Turn RL (ICLR 2026) — self-play on Kuhn poker / TicTacToe transfers to math reasoning
R-Zero		2025.08	Tencent AI Seattle / WashU (Huang et al.)	arXiv 2508.05004	R-Zero: Self-Evolving Reasoning LLM from Zero Data (ICLR 2026) — Challenger and Solver initialised from one model co-evolve; the Challenger targets the Solver's uncertainty frontier
Language Self-Play		2025.09	Meta (Kuba, Gu, …)	arXiv 2509.07414	Language Self-Play For Data-Free Training — a single model alternates between query-generator and responder modes
Reward-Free Self-Evolution		2026.04	Zhang, Ma et al.	arXiv 2604.18131	Training LLM Agents for Spontaneous, Reward-Free Self-Evolution via World Knowledge Exploration
Skill Self-Play		2026.07	Alibaba Qwen	arXiv 2607.22529	Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills — the self-play curriculum is expressed as evolving skills rather than raw problems; bridges this section and 🧩 Skills

📋 Click to view technical details

Resource	Task source	Reward source	What is modified	Loop closure
TTRL	unlabeled test set	majority vote	weights	L2
Absolute Zero	self-proposed code tasks	Python executor	weights (proposer + solver)	L3
R-Zero	Challenger model	Solver self-consistency + Challenger uncertainty reward	weights ×2	L3
SPIRAL	zero-sum games	game outcome	weights	L3
SEAL	given task data	downstream eval after self-edit	weights via self-written SFT	L2 (RL outer loop, fixed eval)
Language Self-Play	self-generated queries	self-judge	weights	L3

📝 Strictness notes

• TTRL improves on a fixed test distribution; it is adaptation, not open-ended growth.
• Absolute Zero / R-Zero / SPIRAL are the cleanest S1+S2+S3 examples at the weight level, but all rely on an external executor or game engine for ground truth — the loop is closed over tasks, not over the verifier.
• SEAL is the canonical "model writes its own weight update" paper; its outer RL loop still uses a human-designed evaluation.

---

🧠 Model — 🔗 Weights + Harness Co-Evolution

Runtime experience becomes both harness artifacts (memory, skills, harness code) and gradient updates.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
Mem-α		2025.09	UCSD / … (Wang, Takanobu, …)	arXiv 2509.25911	Mem-α: Learning Memory Construction via Reinforcement Learning — RL trains the model to decide what to write to memory
MemRL		2026.01	MemTensor (Zhang, Wang, …)	arXiv 2601.03192	MemRL: Self-Evolving Agents via Runtime Reinforcement Learning on Episodic Memory — frozen LLM, RL over which memories to retrieve; the memory is the learnable component
ECHO		2026.01	Li, Jiang et al.	arXiv 2601.06794	No More Stale Feedback: Co-Evolving Critics for Open-World Agent Learning — critic and policy updated together so the critic does not lag the policy's distribution shift
SkillRL		2026.02	UNC (Xia, Chen, …)	arXiv 2602.08234	SkillRL: Evolving Agents via Recursive Skill-Augmented RL — trajectories distilled into a hierarchical SkillBank that is refined recursively from verification failures during RL
OpenClaw-RL		2026.03	Gen-Verse (Wang, Chen, …)	arXiv 2603.10165	OpenClaw-RL: Train Any Agent Simply by Talking — conversational feedback turned into RL signal for a deployed agent
RewardHarness		2026.05	Zhang, Du et al.	arXiv 2605.08703	RewardHarness: Self-Evolving Agentic Post-Training — the reward-computing harness evolves alongside the post-trained policy
Evolving-RL		2026.05	Xiaohongshu / PKU (Fan, Jin, …)	arXiv 2605.10663	Evolving-RL: End-to-End Optimization of Experience-Driven Self-Evolving Capability within Agents — one shared policy acts as experience extractor and solver; GRPO rewards the downstream transfer gain
Learning, Fast and Slow		2026.05	Tiwari, Sareen et al.	arXiv 2605.12484 · blog	Learning, Fast and Slow: Towards LLMs That Adapt Continually — fast in-context / harness adaptation feeding slow weight consolidation
SIA		2026.05	Hexo Labs (Hebbar et al.)	arXiv 2605.27276	SIA: Self Improving AI with Harness & Weight Updates — joint harness + weight loop
EvoTrainer		2026.06	Alibaba DAMO (Chen, Shi, …)	arXiv 2606.03108	EvoTrainer: Co-Evolving LLM Policies and Training Harnesses for Autonomous Agentic RL — the RL trainer (environments, rewards, curricula) is itself evolved by an agent
SEED		2026.07	Wu, Yang et al.	arXiv 2607.14777	SEED: Self-Evolving On-Policy Distillation for Agentic RL — the teacher is the agent's own privileged-context self
Co-Harness		2026.07	Chen, Xiao et al.	arXiv 2607.22688	Co-Harness: Co-Evolving Harnesses and Model Weights for LLM Agents — harness optimisation produces trajectories that are distilled back into the weights, which enables the next harness round

📝 Strictness notes

• This is the section where the model/harness split breaks down on purpose. Lilian Weng's Harness Engineering for Self-Improvement and the optimisation ladder in her companion list put these at the top rung (L5: harness + weights jointly).
• MemRL / Mem-α keep the LLM frozen and learn the memory policy; they are listed here rather than in 💾 Memory because the update is gradient-based.

---

🧰 Harness — ✍️ Prompt & Context Evolution

Improvement without touching weights or agent code: the prompt, playbook, or context is the thing that learns.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
APE		2022.11	U Toronto / Vector (Zhou et al.)	arXiv 2211.01910	Large Language Models Are Human-Level Prompt Engineers (ICLR 2023) — LLM proposes and scores instructions; the first automatic prompt engineer
Self-Refine		2023.03	CMU / AI2 (Madaan et al.)	arXiv 2303.17651	Self-Refine: Iterative Refinement with Self-Feedback (NeurIPS 2023) — generate → self-feedback → refine, no training; the L1 baseline everything else is measured against
OPRO		2023.09	Google DeepMind (Yang et al.)	arXiv 2309.03409	Large Language Models as Optimizers (ICLR 2024) — the LLM reads the trajectory of (prompt, score) pairs and proposes the next prompt
EvoPrompt		2023.09	Microsoft / Tsinghua (Guo et al.)	arXiv 2309.08532	Connecting LLMs with Evolutionary Algorithms Yields Powerful Prompt Optimizers (ICLR 2024) — GA / DE over prompts with the LLM as mutation operator
Promptbreeder		2023.09	Google DeepMind (Fernando et al.)	arXiv 2309.16797	Promptbreeder: Self-Referential Self-Improvement via Prompt Evolution — the mutation prompts are themselves evolved; the first L3 loop at the prompt level
DSPy		2023.10	Stanford (Khattab et al.)	arXiv 2310.03714	DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines (ICLR 2024) — programs, not prompts; optimisers (BootstrapFewShot, MIPRO, GEPA) compile the pipeline against a metric
TextGrad		2024.06	Stanford (Yuksekgonul et al.)	arXiv 2406.07496	TextGrad: Automatic "Differentiation" via Text — natural-language gradients back-propagated through a compound AI system
Dynamic Cheatsheet		2025.04	Stanford (Suzgun, Yuksekgonul, …)	arXiv 2504.07952	Dynamic Cheatsheet: Test-Time Learning with Adaptive Memory — a persistent, self-curated notes buffer accumulated across test queries
GEPA		2025.07	UC Berkeley / Stanford / Databricks (Agrawal et al.)	arXiv 2507.19457	GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning (ICLR 2026 Oral) — reflect on execution traces in natural language, Pareto-evolve prompts; beats GRPO with up to 35× fewer rollouts
ACE		2025.10	Stanford / SambaNova (Zhang, Hu, …)	arXiv 2510.04618	Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models (ICLR 2026) — generator / reflector / curator maintain a growing structured playbook; avoids "context collapse" from rewriting
MCE		2026.01	PKU (Ye, He, … Song)	arXiv 2601.21557	Meta Context Engineering via Agentic Skill Evolution — evolves the context-engineering procedure itself (the L4 "optimizer of the optimizer" rung)

📝 Strictness notes

• Self-Refine is L1 (nothing persists); it is here as the baseline. Spontaneous Reward Hacking in Iterative Self-Refinement (⚠️) shows what goes wrong when the same model is judge and refiner.
• APE / OPRO / EvoPrompt / DSPy optimisers need a labelled dev set as the score; S1 is satisfied by the proposal side only.
• Promptbreeder / MCE evolve the mutator, so they are genuine L3 at the text level.

---

🧰 Harness — 💾 Memory

Experience is written into a persistent store that changes what the agent does next time.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
Reflexion		2023.03	Northeastern / MIT (Shinn et al.)	arXiv 2303.11366	Reflexion: Language Agents with Verbal Reinforcement Learning (NeurIPS 2023) — verbal self-reflections stored in episodic memory across trials
Generative Agents		2023.04	Stanford / Google (Park et al.)	arXiv 2304.03442	Generative Agents: Interactive Simulacra of Human Behavior — memory stream + reflection + planning; the architecture most agent-memory systems descend from
MemoryBank		2023.05	Zhong, Guo et al.	arXiv 2305.10250	MemoryBank: Enhancing LLMs with Long-Term Memory (AAAI 2024) — Ebbinghaus-style forgetting curve over stored memories
ExpeL		2023.08	Tsinghua (Zhao et al.)	arXiv 2308.10144	ExpeL: LLM Agents Are Experiential Learners (AAAI 2024) — distil cross-task insights from successes and failures without weight updates
MemGPT / Letta		2023.10	UC Berkeley → Letta (Packer et al.)	arXiv 2310.08560	MemGPT: Towards LLMs as Operating Systems — the agent pages its own memory between context and archival storage; now the Letta platform
Agent Workflow Memory		2024.09	CMU (Wang, Mao, … Neubig)	arXiv 2409.07429	Agent Workflow Memory — induce reusable workflows from past trajectories, online, for web agents
A-MEM		2025.02	Rutgers (Xu et al.)	arXiv 2502.12110	A-MEM: Agentic Memory for LLM Agents (NeurIPS 2025) — Zettelkasten-style notes that link and update each other when new memories arrive
Mem0		2025.04	Mem0 (Chhikara et al.)	arXiv 2504.19413	Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory — the most-deployed open memory layer
Agent KB		2025.07	Tang, Qin et al.	arXiv 2507.06229	Agent KB: Leveraging Cross-Domain Experience for Agentic Problem Solving — shared knowledge base across agents and domains
What Deserves Memory		2025.08	—	arXiv 2508.03341	What Deserves Memory: Adaptive Memory Distillation for LLM Agents — learn what to keep from experience
Memento		2025.08	UCL / Huawei (Zhou, Chen, … Wang)	arXiv 2508.16153 · Memento 2	Memento: Fine-tuning LLM Agents without Fine-tuning LLMs — case-based reasoning over a memory of (state, action, reward); Memento 2 adds stateful reflective memory
SEDM		2025.09	Xu, Hu et al.	arXiv 2509.09498	SEDM: Scalable Self-Evolving Distributed Memory for Agents — memories carry verified utility and are shared across agents
MemGen		2025.09	Zhang, Fu et al.	arXiv 2509.24704	MemGen: Weaving Generative Latent Memory for Self-Evolving Agents — memory as generated latent tokens woven into reasoning
ReasoningBank		2025.09	Google Cloud AI / UIUC (Ouyang, Yan, …)	arXiv 2509.25140	ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory — distil generalisable reasoning strategies from both successes and self-judged failures; memory-aware test-time scaling
MUSE		2025.10	KnowledgeXLab (Yang et al.)	arXiv 2510.08002	Learning on the Job: An Experience-Driven Self-Evolving Agent for Long-Horizon Tasks — hierarchical memory updated after every sub-task
EvolveR		2025.10	KnowledgeXLab (Wu, Wang, …)	arXiv 2510.16079	EvolveR: Self-Evolving LLM Agents through an Experience-Driven Lifecycle — offline distillation of principles + online RL-with-retrieval
WebCoach		2025.11	—	arXiv 2511.12997	WebCoach: Self-Evolving Web Agents with Cross-Session Memory Guidance
ReMe		2025.12	Alibaba AgentScope (Cao, Deng, …)	arXiv 2512.10696	Remember Me, Refine Me: A Dynamic Procedural Memory Framework for Experience-Driven Agent Evolution
MemEvolve		2025.12	Zhang, Ren et al.	arXiv 2512.18746	MemEvolve: Meta-Evolution of Agent Memory Systems — evolves the memory architecture, not just its contents (L4)
Agentic Memory		2026.01	Yu, Yao et al.	arXiv 2601.01885	Learning Unified Long-Term and Short-Term Memory Management for LLM Agents
Live-Evo		2026.02	—	arXiv 2602.02369	Live-Evo: Online Evolution of Agentic Memory from Continuous Feedback
MemSkill		2026.02	Zhang, Long et al.	arXiv 2602.02474	MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents — memory operations are skills that themselves evolve
SAGE		2026.05	Wang, Zhao et al.	arXiv 2605.12061	SAGE: A Self-Evolving Agentic Graph-Memory Engine for Structure-Aware Associative Memory
CASCADE		2026.05	KCL (Guo, Du, …)	arXiv 2605.06702	CASCADE: Case-Based Continual Adaptation for LLMs During Deployment — plus the DTLBench deployment-time-learning benchmark
Faulty Memories		2026.05	—	arXiv 2605.12978	Useful Memories Become Faulty When Continuously Updated by LLMs — negative result: drift under continual LLM rewriting
AutoMem		2026.07	Wu, Zhu et al.	arXiv 2607.01224	AutoMem: Automated Learning of Memory as a Cognitive Skill

📝 Strictness notes

• Reflexion / Generative Agents / MemGPT persist per-task or per-session state; whether they are "self-improving" depends on whether memory survives across tasks. ExpeL, AWM and ReasoningBank are the ones that explicitly make cross-task, self-judged insights.
• For the full memory landscape (products, benchmarks, parametric memory) see TeleAI-UAGI/Awesome-Agent-Memory; this section keeps only memory systems that evolve from the agent's own experience.
• Faulty Memories (2605.12978) is the caution for the whole section.

---

🧰 Harness — 🧩 Skills

Procedural knowledge packaged as reusable, versioned skills that the agent discovers, tests, and refines — the layer that Claude Code / Codex "agent skills" made mainstream in 2026.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
Voyager		2023.05	NVIDIA / Caltech (Wang et al.)	arXiv 2305.16291	Voyager: An Open-Ended Embodied Agent with LLMs (Seminal) — automatic curriculum + an ever-growing skill library of verified code + iterative prompting; the origin of "skills" as the unit of self-improvement
CRADLE		2024.03	BAAI et al. (Tan et al.)	arXiv 2403.03186	Cradle: Empowering Foundation Agents Towards General Computer Control — Voyager-style skill curation (tool creation + knowledge discovery) for general computer control
SkillWeaver		2025.04	Ohio State (Zheng, Fatemi, … Su)	arXiv 2504.07079	SkillWeaver: Web Agents can Self-Improve by Discovering and Honing Skills — propose → practice → synthesise APIs → self-test; skills transfer between weak and strong agents
Agent Skills		2025.10	Anthropic	repo	Agent Skills — the SKILL.md format (instructions + scripts + resources, progressively loaded) that made skills a portable harness primitive across Claude Code, Codex, Cursor, …
Superpowers		2025.10	Jesse Vincent	repo	Superpowers — an agentic skills framework + methodology (brainstorm → plan → TDD → review) with a writing-skills skill, i.e. skills that create skills
Skill-Pro		2026.02	Mi, Ma et al.	arXiv 2602.01869	Skill-Pro: Learning Reusable Skills from Experience via Non-Parametric PPO for LLM Agents (a.k.a. ProcMEM) — PPO-style updates on a textual skill store instead of weights
SkillsBench		2026.02	—	arXiv 2602.12670	SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks — do skills help? measured
AutoSkill		2026.03	Yang, Li et al.	arXiv 2603.01145	AutoSkill: Experience-Driven Lifelong Learning via Skill Self-Evolution
EvoSkill		2026.03	—	arXiv 2603.02766	EvoSkill: Automated Skill Discovery for Multi-Agent Systems
XSkill		2026.03	—	arXiv 2603.12056	XSkill: Continual Learning from Experience and Skills in Multimodal Agents
SWE-Skills-Bench		2026.03	—	arXiv 2603.15401	SWE-Skills-Bench: Do Agent Skills Actually Help in Real-World Software Engineering? — sober measurement of skill benefit on SWE tasks
Memento-Skills		2026.03	UCL (Zhou, Guo, …)	arXiv 2603.18743	Memento-Skills: Let Agents Design Agents — a meta-agent writes the skills that define new agents
Trace2Skill		2026.03	—	arXiv 2603.25158	Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills
CoEvoSkills		2026.04	—	arXiv 2604.01687	CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification — skills and their verifiers co-evolve
SKILL0		2026.04	—	arXiv 2604.02268	SKILL0: In-Context Agentic RL for Skill Internalization — skills move from context into weights
SkillForge		2026.04	Liu, Luo et al.	arXiv 2604.08618	SkillForge: Forging Domain-Specific, Self-Evolving Agent Skills in Cloud Technical Support — production deployment report
SkillEvolver		2026.05	—	arXiv 2605.10500	SkillEvolver: Skill Learning as a Meta-Skill — the skill-learning procedure is itself a skill (L4)
SkillOpt		2026.05	Microsoft (Yang, Gong, …)	arXiv 2605.23904	SkillOpt: Executive Strategy for Self-Evolving Agent Skills
CODESKILL		2026.05	Li, Zhang et al.	arXiv 2605.25430	CODESKILL: Learning Self-Evolving Skills for Coding Agents — managed procedural skills lift coding pass rate 29.6 → 39.3 vs no-skill baseline
MUSE-Autoskill		2026.05	—	arXiv 2605.27366	MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation
OpenSkill		2026.06	Yan, Song et al.	arXiv 2606.06741	OpenSkill: Open-World Self-Evolution for LLM Agents
Co-Evolving Skill Gen		2026.06	Zhang, Lin et al.	arXiv 2606.08755	Co-Evolving Skill Generation and Policy Optimization
Skill Eval & Evolution		2026.06	Ding, Zhou et al.	arXiv 2606.11435	Agent Skill Evaluation and Evolution: Frameworks and Benchmarks
SkillProx		2026.08	Zheng, Zhou et al.	arXiv 2608.07449	SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient Descent — proximal updates in text space; deletion is a first-class operation
ERSkill		2026.08	Chen, Zhang et al.	arXiv 2608.12720	ERSkill: Evolving for Skill-Guided Adaptive Memory Retrieval — the retrieval behaviour itself becomes a skill
SkillCommit		2026.08	He, Yang et al.	arXiv 2608.15165	SkillCommit: Evolving Agent Skills through Behaviorally Validated Scope Expansion — against semantic-similarity merging; commit hierarchical abstractions only after behavioural validation
HyperSkill		2026.08	Xu, Yang et al.	arXiv 2608.16114	HyperSkill: Self-Evolving LLM Agents via Hypergraph-Structured Skill Memory
WikiSkill		2026.08	Google / Virginia Tech (Tang, Rashtchian, … Vu)	arXiv 2608.27454	WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution — experience → structured wiki → distilled skills, so knowledge compounds across tasks
Hivemind		2026.04	Activeloop	repo	Hivemind — continual-learning layer that distils coding-agent session trajectories into reusable skills

📋 Click to view technical details

Resource	Skill representation	How skills are created	How skills are validated	Where stored
Voyager	executable JS code + description	LLM writes from curriculum task	environment execution + self-verification	vector-indexed library
SkillWeaver	Python API	practice + synthesis	self-test on the website	per-site library
Anthropic Agent Skills	SKILL.md + scripts	human or agent authored	none built-in	filesystem, progressive disclosure
Skill-Pro / SkillProx	text	PPO / proximal text-gradient updates	task reward	textual store
CoEvoSkills	text + verifier	co-evolution	evolved verifier	library
WikiSkill	wiki pages → skills	compile experience	downstream success	persistent wiki
SkillCommit	hierarchical abstractions	scope expansion	behavioural validation before commit	versioned

📝 Strictness notes

• Anthropic Agent Skills / Superpowers are formats and frameworks, not learning methods; they are listed because they define the artifact that the 2026 papers evolve. Skills that agents author for themselves satisfy S2; whether S1/S3 hold depends on the surrounding loop.
• SkillsBench / SWE-Skills-Bench are negative-to-mixed results on whether human-written skills help at all — read before assuming skill libraries are free wins.
• Aug 2026 produced five parallel skill-evolution papers (SkillProx, ERSkill, SkillCommit, HyperSkill, WikiSkill); they disagree mainly on how skills are merged and retired.

---

🧰 Harness — 🕸️ Workflow & Agent-Architecture Search

The topology of the agent (roles, edges, control flow) is the thing being optimised.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
Agent Symbolic Learning		2024.06	AIWaves (Zhou et al.)	arXiv 2406.18532	Symbolic Learning Enables Self-Evolving Agents — "language gradients" over prompts, tools and pipeline, back-propagated symbolically
ADAS		2024.08	UBC / Vector (Hu, Lu, Clune)	arXiv 2408.08435	Automated Design of Agentic Systems (ICLR 2025) — Meta Agent Search: a meta-agent programs new agents in code, keeps an archive of discoveries
AgentSquare		2024.10	Tsinghua FIB (Shang et al.)	arXiv 2410.06153	AgentSquare: Automatic LLM Agent Search in Modular Design Space (ICLR 2025) — planning / reasoning / tool / memory modules recombined by evolution
AFlow		2024.10	FoundationAgents / MetaGPT (Zhang et al.)	arXiv 2410.10762	AFlow: Automating Agentic Workflow Generation (ICLR 2025 Oral) — MCTS over code-represented workflows
Agentic Supernet		2025.02	Zhang, Niu et al.	arXiv 2502.04180	Multi-agent Architecture Search via Agentic Supernet — NAS-style supernet over multi-agent systems
Alita		2025.05	Princeton (Qiu et al.)	arXiv 2505.20286	Alita: Generalist Agent Enabling Scalable Agentic Reasoning with Minimal Predefinition and Maximal Self-Evolution — the agent creates its own MCP tools on demand
EvoAgentX		2025.07	Glasgow / EvoAgentX (Wang et al.)	arXiv 2507.03616	EvoAgentX: An Automated Framework for Evolving Agentic Workflows — open framework unifying TextGrad / AFlow / MIPRO-style optimisers
Group-Evolving Agents		2026.02	Weng, Antoniades et al.	arXiv 2602.04837	Group-Evolving Agents: Open-Ended Self-Improvement via Experience Sharing — a population of agents shares experience instead of a single lineage
CORAL		2026.04	Qu, Zheng et al.	arXiv 2604.01658	CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery

---

🧰 Harness — 🔧 Self-Modifying Harness Code

The agent rewrites its own scaffold: tools, control loop, prompts-as-code, or the whole repository. The Gödel Machine made empirical.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
STOP		2023.10	Stanford / Microsoft (Zelikman, Lorch, Mackey, Kalai)	arXiv 2310.02304	Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation (COLM 2024) — a scaffold that improves code is applied to itself, with a frozen model; the first empirical L3/L4
Gödel Agent		2024.10	PKU (Yin et al.)	arXiv 2410.04444	Gödel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement (ACL 2025) — monkey-patches its own runtime logic guided only by a high-level objective
SICA		2025.04	Bristol / iGent (Robeyns, Szummer, Aitchison)	arXiv 2504.15228	A Self-Improving Coding Agent — the agent edits its own codebase against a utility of benchmark score, cost and time; no gradients
DGM		2025.05	UBC / Vector / Sakana (Zhang, Hu, Lu, Lange, Clune)	arXiv 2505.22954 · blog	Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents (ICLR 2026) — an archive of coding agents that self-modify; empirical validation on SWE-bench / Polyglot replaces proofs; 20 → 50% SWE-bench
HGM		2025.10	KAUST / MetaAUTO (Wang, Piękos, … Schmidhuber)	arXiv 2510.21614	Huxley-Gödel Machine — estimates clade-level productivity (how good an agent's descendants are) instead of the agent's own score, to pick what to expand
Live-SWE-agent		2025.11	UIUC (Xia, Wang, …)	arXiv 2511.13646	Live-SWE-agent: Can Software Engineering Agents Self-Evolve on the Fly? — builds its own tools during the task
DARWIN		2026.02	Henry Jiang	arXiv 2602.05848	DARWIN: Dynamic Agentically Rewriting Self-Improving Network
HyperAgents		2026.03	UBC / Oxford / Microsoft (Zhang, Zhao, … Clune, Jiang, Devlin, Shavrina)	arXiv 2603.19461	Hyperagents — the meta-agent that modifies agents is itself modifiable; the DGM lineage's answer to "who improves the improver"
Meta-Harness		2026.03	Stanford IRIS (Lee, Nair, Zhang, Lee, …)	arXiv 2603.28052	Meta-Harness: End-to-End Optimization of Model Harnesses — treats the whole harness as the optimisation variable on Terminal-Bench 2
AHE		2026.04	Fudan et al. (Lin, Liu, …)	arXiv 2604.25850	Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses — traces → diagnoses → harness patches
Continual Harness		2026.05	Princeton (Karten, Zhang, … Jin, Vodrahalli)	arXiv 2605.09998	Continual Harness: Online Adaptation for Self-Improving Foundation Agents — reset-free: fix the harness at the point of failure instead of restarting
MOSS		2026.05	HKGAI (Cai, Zhang, … Guo)	arXiv 2605.22794	MOSS: Self-Evolution through Source-Level Rewriting in Autonomous Agent Systems — production-oriented source rewriting with versioning
DemoEvolve		2026.05	Che, Yang et al.	arXiv 2605.24539	DemoEvolve: Overcoming Sparse Feedback in Agentic Harness Evolution with Demonstrations
Harness Updating ≠ Benefit		2026.05	Lin, Wu et al.	arXiv 2605.30621	Harness Updating Is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents — separates "can edit itself" from "edits help"; a needed negative result
Adaptive Auto-Harness		2026.06	Liu, Shi et al.	arXiv 2606.01770	Adaptive Auto-Harness: Sustained Self-Improvement for Agentic System Deployment on Open-Ended Task Streams — reports early-peak-then-decay of dense self-improvement on open streams
RHO		2026.06	Pan, Liu et al.	arXiv 2606.05922	Evolving Agents in the Dark: Retrospective Harness Optimization via Self-Preference — fully label-free; SWE-Bench Pro 59 → 78%
HarnessFix		2026.06	Chen, Wang et al.	arXiv 2606.06324	From Failed Trajectories to Reliable LLM Agents: Diagnosing and Repairing Harness Flaws — compiles trajectory + harness to an IR and attributes failures to seven layers
Self-Harness		2026.06	Shanghai AI Lab (Zhang, Zhang, … Bai, Hu)	arXiv 2606.09498	Self-Harness: Harnesses That Improve Themselves
Held-Out Selection		2026.06	Nguyen, Nguyen	arXiv 2606.28374	Recursive Self-Evolving Agents via Held-Out Selection — selection on held-out tasks to stop the loop from overfitting its own evaluator
Self-Evolving Coding Agents		2026.08	Zhou, Hu et al.	arXiv 2608.03392	Self-Evolving Coding Agents — survey/position: agents that update framework, memory, skills, tools, models or collaboration structure from prior coding interactions
Ouroboros		2026.08	Razzhigaev, Gritsaev et al.	arXiv 2608.08311	Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution — core changes pass a review gate before merge
HSI		2026.08	Tailin Zhou	arXiv 2608.08466	Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses — three scopes of evolution with a frozen meta-evolver as the outer anchor
Evo-Harness		2026.08	Wei, Shi et al.	arXiv 2608.15071	Evo-Harness: Context-to-Harness Skill Compilation for Self-Evolving Agents — reflections compiled into harness-level skills, variables isolated across five benchmarks
EnvHarness		2026.08	Google Cloud AI (Huang, Wang, …)	arXiv 2608.19880	EnvHarness: Awakening Static Worlds for Agent Learning — evolve the environment side of the harness
AutoSaddler		2026.08	POSTECH / KAIST / Microsoft (Park, Kim, …)	arXiv 2608.23041	AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces
Prime Agent		2026.08	Princeton / Prime Intellect / MIT (Karten, Zhang, …)	arXiv 2608.23552	Prime Agent: A Self-Improving RLM Harness — open harness for recursive language models that improves itself from its own runs
Metaⁿ		2026.08	Kim, Lee et al.	arXiv 2608.24735	Metaⁿ: Recursive Self-Improvement through Emergent Depth — argues useful self-rewrite meta-depth saturates around 2.5 levels; change the input, not the machine, to go deeper

📋 Click to view technical details

Resource	What is rewritten	Selection signal	Archive / population	Loop closure
STOP	the improver scaffold	utility on downstream tasks	none (single lineage)	L4
SICA	own repo	benchmark score − cost − time	single agent	L3
DGM	own repo (tools, workflow)	SWE-bench / Polyglot	open-ended archive	L3
HGM	own repo	clade productivity estimate	tree	L3
HyperAgents	agent and meta-agent	task benchmarks	archive	L4
Meta-Harness	full harness config/code	Terminal-Bench 2	—	L3
RHO	harness	self-preference (no labels)	—	L3
Ouroboros	core code	reviewed merge gate	git history	L3 with human/AI review

📝 Strictness notes

• Every system here still uses a fixed external benchmark as the selection signal; the only things that evolve the evaluator are in ⚖️ below. Held-Out Selection and Harness Updating ≠ Benefit are the two papers to read before believing a self-modification curve.
• DGM and SICA run with sandboxing and human-inspectable diffs; the papers explicitly document reward-hacking incidents (e.g. faking test logs).
• The harness-engineering practice pieces (OpenAI, Anthropic, Fowler, Osmani, Weng) are in 📰 Blogs.

---

🧰 Harness — ⚖️ Evolving Evaluators

Who grades the grader: rubrics, judges and metrics that improve together with the agent they judge.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
CycleResearcher		2024.10	Westlake (Weng, Zhu, …)	arXiv 2411.00816	CycleResearcher: Improving Automated Research via Automated Review (ICLR 2025) — a trained reviewer model closes the loop on a trained paper-writer
Red Queen Gödel Machine		2026.06	Cambridge et al. (Iacob, Jovanović, Shen, …)	arXiv 2606.26294	The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators — DGM where the benchmark also evolves
SCORE		2026.06	Zhu, Cai et al.	arXiv 2606.04507	Self-Evolving Deep Research via Joint Generation and Evaluation — generator and evaluator share parameters and train jointly
Who Grades the Grader?		2026.07	Zhang, Wang et al.	arXiv 2607.12790	Co-Evolving Evaluation Metrics and Skills for Self-Improving LLM Agents
EvalCEGAR		2026.08	Zhang, Cui et al.	arXiv 2608.18744	Metrics That Write Themselves: Evolving an Evaluator from Its Own Blind Spots — counterexample-guided (CEGAR-style) evolution of the evaluator
EvoLM		2026.05	UW	arXiv 2605.03871	co-evolved rubrics at the weight level (cross-listed from 🧪)

📝 Strictness notes

• This is the most fragile part of the stack: once the evaluator moves, "improvement" can be an artifact. Held-Out Selection (🔧) and HVTB (📊) are the counter-measures; Falsifiable Release Gates (⚠️) is the governance answer.

---

🏗️ Infra — ⚙️ Kernels

AI writing and optimising the GPU / NPU kernels that AI runs on. The tightest infra loop: the reward is a profiler.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
KernelBench		2025.02	Stanford (Ouyang, Guo, … Mirhoseini)	arXiv 2502.10517 · blog	KernelBench: Can LLMs Write Efficient GPU Kernels? — 250 PyTorch → CUDA tasks, fast_p metric; the benchmark every kernel agent reports on
AI CUDA Engineer		2025.02	Sakana AI	report · robust-kbench	The AI CUDA Engineer: Agentic CUDA Kernel Discovery, Optimization and Composition — evolutionary kernel optimisation with a retrieval archive of past kernels; also the famous reward-hacking incident (kernels that bypassed the correctness check), fixed in Towards Robust Agentic CUDA Kernel Benchmarking
DeepSeek-R1 kernel gen		2025.02	NVIDIA	blog	Automating GPU Kernel Generation with DeepSeek-R1 and Inference-Time Scaling — verifier-in-the-loop generation of attention kernels
KernelLLM		2025.06	Meta	HF	KernelLLM — 8B model SFT'd on PyTorch → Triton pairs
AutoTriton		2025.07	Tsinghua (THUNLP)	arXiv 2507.05687	AutoTriton: Automatic Triton Programming with RL in LLMs
Kevin		2025.07	Cognition	arXiv 2507.11948	Kevin: Multi-Turn RL for Generating CUDA Kernels — multi-turn RL with compiler/profiler feedback in the loop
CUDA-L1		2025.07	DeepReinforce (Li, Wang, …)	arXiv 2507.14111 · CUDA-L2	CUDA-L1: Improving CUDA Optimization via Contrastive RL — contrastive RL on speedup; CUDA-L2 (Dec 2025) reports beating cuBLAS on matmul
GEAK		2025.07	AMD (Wang, Joshi, …)	arXiv 2507.23194	Geak: Introducing Triton Kernel AI Agent & Evaluation Benchmarks — Triton agent + benchmarks for AMD GPUs
Astra		2025.09	Stanford (Wei, Sun, …)	arXiv 2509.07506	Astra: A Multi-Agent System for GPU Kernel Performance Optimization
EvoEngineer		2025.10	Guo, Zhu et al.	arXiv 2510.03760	EvoEngineer: Mastering Automated CUDA Kernel Code Evolution with LLMs
TritonRL		2025.10	—	arXiv 2510.17891	TritonRL: Training LLMs to Think and Code Triton Without Cheating — reward design against verifier gaming
FM Agent		2025.10	Li, Wu et al.	arXiv 2510.26144	The FM Agent — general evolutionary agent applied to kernels among other domains
CudaForge		2025.10	Zhang, Wang et al.	arXiv 2511.01884	CudaForge: An Agent Framework with Hardware Feedback for CUDA Kernel Optimization
PRAGMA		2025.11	Lei, Yang et al.	arXiv 2511.06345	PRAGMA: A Profiling-Reasoned Multi-Agent Framework for Automatic Kernel Optimization
KernelFalcon		2025.11	Meta / PyTorch	blog · BackendBench	KernelFalcon: Autonomous GPU Kernel Generation via Deep Agents — 100% on KernelBench L1–L3 with a hierarchical deep-agent harness
KernelBand		2025.11	Ran, Xie et al.	arXiv 2511.18868	KernelBand: Steering LLM-based Kernel Optimization via Hardware-Aware Multi-Armed Bandits
AccelOpt		2025.11	Stanford / AWS (Zhang, Zhu, …)	arXiv 2511.15915	AccelOpt: A Self-Improving LLM Agentic System for AI Accelerator Kernel Optimization (MLSys 2026) — an optimisation memory of past successes/failures makes later Trainium kernels better; explicitly self-improving
QiMeng-Kernel		2025.11	ICT CAS (QiMeng)	arXiv 2511.20100	QiMeng-Kernel: Macro-Thinking Micro-Coding Paradigm for LLM-Based High-Performance GPU Kernel Generation
TritonForge		2025.12	Li, Man et al.	arXiv 2512.09196	TritonForge: Profiling-Guided Framework for Automated Triton Kernel Optimization
cuPilot		2025.12	Chen, Wu et al.	arXiv 2512.16465	cuPilot: A Strategy-Coordinated Multi-agent Framework for CUDA Kernel Evolution
KernelEvolve		2025.12	Meta (Liao, Qin, …)	arXiv 2512.23236 · Meta blog	KernelEvolve: Scaling Agentic Kernel Coding for Heterogeneous AI Accelerators at Meta — production: NVIDIA / AMD / MTIA / CPU kernels; a knowledge base grows with every run; +60% ads-model inference throughput in hours
AKG Agent		2025.12	Huawei MindSpore	arXiv 2512.23424	AKG kernel Agent: A Multi-Agent Framework for Cross-Platform Kernel Synthesis
FlashInfer-Bench		2026.01	FlashInfer / CMU / NVIDIA (Xing, Zhai, …)	arXiv 2601.00227 · blog	FlashInfer-Bench: Building the Virtuous Cycle for AI-driven LLM Systems — agent-generated kernels are deployed into the serving engine, whose traces become the next benchmark: the self-improving serving loop
Dr. Kernel / KernelGYM		2026.02	HKUST (Liu, Xu, …)	arXiv 2602.05885	Dr. Kernel: Reinforcement Learning Done Right for Triton Kernel Generations (ICML 2026) — distributed GPU RL environment + recipe
KernelBlaster		2026.02	Dong, Modi et al.	arXiv 2602.14293	KernelBlaster: Continual Cross-Task CUDA Optimization via Memory-Augmented In-Context RL
K-Search		2026.02	UC Berkeley (Cao, Mao, …)	arXiv 2602.19128	K-Search: LLM Kernel Generation via Co-Evolving Intrinsic World Model — the agent learns a performance world model of the GPU while searching
CUDA Agent		2026.02	ByteDance Seed / Tsinghua (Dai, Wu, …)	arXiv 2602.24286	CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation
KernelSkill		2026.03	Sun, Han et al.	arXiv 2603.10085	KernelSkill: A Multi-Agent Framework for GPU Kernel Optimization — kernel optimisation knowledge as reusable skills
Value-Driven Memory (NPU)		2026.03	Zheng, Li et al.	arXiv 2603.10846	Towards Cold-Start Drafting and Continual Refining: A Value-Driven Memory Approach with Application to NPU Kernel Synthesis
KernelFoundry		2026.03	Wiedemann, Leboutet et al.	arXiv 2603.12440	KernelFoundry: Hardware-aware evolutionary GPU kernel optimization
AutoKernel		2026.03	RightNow AI (Jaber et al.)	arXiv 2603.21331	AutoKernel: Autonomous GPU Kernel Optimization via Iterative Agent-Driven Search — "autoresearch for GPU kernels": give it a PyTorch model, wake up to Triton kernels
Kernel-Smith		2026.03	Du, Ge et al.	arXiv 2603.28342	Kernel-Smith: A Unified Recipe for Evolutionary Kernel Optimization
AdaExplore		2026.04	Du, Zhuo et al.	arXiv 2604.16625	AdaExplore: Failure-Driven Adaptation and Diversity-Preserving Search for Efficient Kernel Generation
ARGUS		2026.04	Mai, Guo et al.	arXiv 2604.18616	ARGUS: Agentic GPU Optimization Guided by Data-Flow Invariants
Kernel Design Agents		2026.05	MIT HAN Lab	repo	Kernel Design Agents — open agent harness for kernel design
KernelBenchX		2026.05	Wang, Zhang et al.	arXiv 2605.04956 · SOL-ExecBench	KernelBenchX: A Comprehensive Benchmark for Evaluating LLM-Generated GPU Kernels — with NVIDIA's SOL-ExecBench (speed-of-light vs hardware limits) the current benchmark pair

📋 Click to view technical details

Resource	Method	Signal	Persistent artifact that improves	Loop closure
AutoTriton / Kevin / CUDA-L1 / Dr. Kernel / CUDA Agent	RL on the LLM	compile + correctness + speedup	model weights	L2
AI CUDA Engineer / EvoEngineer / Kernel-Smith / KernelFoundry	evolutionary search	profiler	kernel archive	L2
AccelOpt / KernelBlaster / KernelSkill / Value-Driven Memory	agent + memory	profiler	optimisation memory / skills	L3 (memory feeds later runs)
KernelEvolve	agent + knowledge base	profiler, production traffic	knowledge base + shipped kernels	L3, in production
FlashInfer-Bench	agent ↔ serving engine	serving traces	deployed kernels + next benchmark	L3
K-Search	search + learned world model	profiler	world model	L3

📝 Strictness notes

• Pure RL-for-kernels papers improve a model that writes kernels, not the system that trains it (L2). The entries that satisfy S3 are the memory/knowledge-base ones (AccelOpt, KernelEvolve, KernelBlaster, KernelSkill) and FlashInfer-Bench's deploy-and-remeasure loop.
• The section's founding failure is the AI CUDA Engineer's correctness-check bypass; TritonRL, robust-kbench and SOL-ExecBench exist because of it.
• For the complete kernel-agent landscape (60+ entries, datasets, DSLs) see flagos-ai/awesome-LLM-driven-kernel-generation.

---

🏗️ Infra — 🧮 Compilers, Serving & Configuration

AI tuning the compiler passes, serving configs, schedulers and chips underneath the model.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
Chip Placement		2020.04	Google (Mirhoseini, Goldie, …)	arXiv 2004.10746	Chip Placement with Deep Reinforcement Learning (Nature 2021, "AlphaChip") — RL places the floorplan of the TPUs that train the next models; the earliest production "AI designs AI hardware" loop
MLGO		2021.01	Google (Trofin et al.)	arXiv 2101.04808	MLGO: a Machine Learning Guided Compiler Optimizations Framework — learned inlining / register allocation shipped in LLVM
LLM Compiler		2024.07	Meta (Cummins et al.)	arXiv 2407.02524	Meta Large Language Model Compiler: Foundation Models of Compiler Optimization — LLMs trained on IR and assembly to predict optimal pass sequences
AlphaEvolve (infra results)		2025.05	Google DeepMind	paper · blog	AlphaEvolve applied to Google's own stack — a Borg scheduling heuristic recovering 0.7% of fleet compute, a 23% faster Gemini matmul kernel (1% less Gemini training time), a Verilog simplification adopted in a TPU, 32.5% faster FlashAttention: the clearest public case of a model improving the infra that trains it
AIConfigurator		2026.01	NVIDIA (Xu, Liu, …)	arXiv 2601.06288	AIConfigurator: Lightning-Fast Configuration Optimization for Multi-Framework LLM Serving — operation-level performance model to pick TP/PP/batch/KV settings for vLLM / SGLang / TRT-LLM / Dynamo in seconds
ISO-Bench		2026.02	Lossfunk (Nangia, Mishra, …)	arXiv 2602.19594	ISO-Bench: Can Coding Agents Optimize Real-World Inference Workloads? — agents vs real vLLM / SGLang performance PRs
AutoPass		2026.06	Li, Ren et al.	arXiv 2606.20373	AutoPass: Evidence-Guided LLM Agents for Compiler Performance Tuning — inference-only agents tune pass pipelines with profiling evidence
Tool-Making in Low-Latency Systems		2026.07	Kujanpää, Liu et al.	arXiv 2607.08010	Tool-Making and Self-Evolving LLM Agents in Low-Latency Systems — repeated SOP steps compiled into validated, versioned tools before deployment

📝 Strictness notes

• AlphaChip / MLGO / LLM Compiler are learned optimisers inside the toolchain; they become self-improvement only when the models they help train are the ones proposing the next optimisation (which AlphaEvolve's Gemini-kernel result is the first public example of).
• AIConfigurator is an analytical tuner, not an agent; listed because it is the config-search primitive an infra agent would call.

---

🏗️ Infra — 🏭 Training Pipelines & Data Engines

Agents that run the training loop: edit the training script, pick hyper-parameters, build the dataset, post-train the model.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
AIDE		2025.02	Weco AI (Jiang, Schmidt, …)	arXiv 2502.13138	AIDE: AI-Driven Exploration in the Space of Code — tree search over ML solution scripts; the reference agent in OpenAI's MLE-bench
ML-Master		2025.06	SJTU (Liu, Cai, …)	arXiv 2506.16499	ML-Master: Towards AI-for-AI via Integration of Exploration and Reasoning — MLE agent with adaptive memory of its own exploration
AIRA		2025.07	Meta FAIR / UCL (Toledo, Hambardzumyan, …)	arXiv 2507.02554	AI Research Agents for Machine Learning: Search, Exploration, and Generalization in MLE-bench — separates search policy from operator set; generalisation gap between validation and test is the bottleneck
Adaptive Data Flywheel		2025.10	NVIDIA (Shukla, Knowles, …)	arXiv 2510.27051	Adaptive Data Flywheel: Applying MAPE Control Loops to AI Agent Improvement — monitor → analyse → plan → execute over a deployed agent's data
autoresearch		2026.03	Andrej Karpathy	repo · tweet · AINews	autoresearch — a 630-line single-GPU nanochat training script plus a recipe: the agent edits train.py, runs a fixed 5-minute experiment, keeps or reverts, repeats. ~700 experiments in two days found ~20 stacking improvements (≈11% faster GPT-2-scale training). The "Karpathy loop" that put self-improving training pipelines in the mainstream; Karpathy joined Anthropic in May 2026 to run it inside Claude pretraining
PostTrainBench		2026.03	AISA (Rank, Bhatnagar, …)	arXiv 2603.08640	PostTrainBench: Can LLM Agents Automate LLM Post-Training? — Claude Code / Codex CLI given a base model, one H100 and 10 hours
AiScientist (long-horizon MLE)		2026.04	AweAI (Chen, Chen, …)	arXiv 2604.13018	Toward Autonomous Long-Horizon Engineering for ML Research
AgenticQwen		2026.04	Alibaba (Lyu, Wang, …)	arXiv 2604.21590	AgenticQwen: Training Small Agentic LMs with Dual Data Flywheels for Industrial-Scale Tool Use — reasoning and agentic flywheels with strong-model validation before training
GEAR		2026.05	Vector / U Toronto (Jeddi, Le, …)	arXiv 2605.13874	GEAR: Genetic AutoResearch for Agentic Code Evolution — population-based autoresearch
MLReplicate		2026.05	Gaddipati, Muhammed et al.	arXiv 2605.16616	MLReplicate: Benchmarking Autonomous Research Systems for ML Reproducibility
MLEvolve		2026.06	Shanghai AI Lab InternScience (Du, Yan, …)	arXiv 2606.06473	MLEvolve: A Self-Evolving Framework for Automated ML Algorithm Discovery
Autodata		2026.06	Meta FAIR	arXiv 2606.25996	agentic synthetic-data scientist (cross-listed from 🧪)
AIDE²		2026.07	Weco AI	blog · 4 levels of RSI	AIDE²: First Evidence of Recursive Self-Improvement — an outer agent rewrites the inner AIDE research agent's code and strategy; in 8 days it found a better autoresearch harness than two years of human tuning (novel search algorithm, 16× smaller prompt, layered anti-reward-hacking). Self-reported, on a held-out benchmark
iCoder		2026.08	SJTU / NUS / DP Technology (Yang, Lyu, …)	report	iCoder: Recursive AI-Led Development of Frontier Industrial Coding Model — a 27B coding model whose data, training and release were driven by AI with humans reduced to a low-frequency gate
Large Discovery Models		2026.08	Yu, Song et al.	arXiv 2608.15669 · project	Large Discovery Models: Empirically-grounded Model-Based Open-Ended Search

📋 Click to view technical details

Resource	What the agent edits	Evaluation	Time per experiment	Loop closure
autoresearch	train.py of a small GPT	val loss after fixed 5-min budget	5 min	L2 (inner model ≠ agent)
AIDE	ML solution script	Kaggle-style metric	minutes–hours	L2
AIDE²	AIDE's own code + strategy	held-out research benchmark	days	L3 (outer agent improves inner agent)
PostTrainBench	post-training recipe	downstream evals	10 h	L2
Autodata / AgenticQwen	training data	strong-model validation, downstream	—	L2
KernelEvolve / FlashInfer-Bench (⚙️)	kernels in production	throughput	hours	L3

📝 Strictness notes

• autoresearch is often called RSI; strictly it is not: the agent improves the training of a different, much smaller model, not its own. It is in this list because it is the recipe everyone now copies (AutoKernel, GEAR, AIDE²).
• AIDE² and iCoder are the two 2026 claims that come closest to S3 on the infra axis; both are self-reported and narrow.

---

🏗️ Infra — 🧬 Program & Algorithm Discovery

Evolutionary search over programs that produces better algorithms, rewards, environments — and increasingly, better versions of the search itself.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
POET		2019.01	Uber AI (Wang, Lehman, Clune, Stanley)	arXiv 1901.01753 · Enhanced POET	Paired Open-Ended Trailblazer — co-evolves environments and the agents that solve them; the open-endedness lineage that leads to DGM
ELM		2022.06	OpenAI / CarperAI (Lehman, Gordon, … Stanley)	arXiv 2206.08896	Evolution through Large Models — LLM as the mutation operator in quality-diversity search; the seed idea of AlphaEvolve
Eureka		2023.10	NVIDIA / UPenn (Ma et al.)	arXiv 2310.12931	Eureka: Human-Level Reward Design via Coding LLMs (ICLR 2024) — GPT-4 evolves reward functions for RL from environment source + training curves
FunSearch		2023.12	Google DeepMind (Romera-Paredes et al.)	Nature	Mathematical Discoveries from Program Search with LLMs (Nature 2023) — LLM + evaluator in an island-based evolutionary loop finds new cap-set constructions and bin-packing heuristics
AlphaEvolve		2025.05	Google DeepMind (Novikov et al.)	arXiv 2506.13131 · blog · impact	AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery — Gemini ensemble evolves whole codebases against automated evaluators; 4×4 complex matmul in 48 multiplications, plus the Google-infra results in 🧮
ShinkaEvolve		2025.09	Sakana AI (Lange, Imajuku, Cetin)	arXiv 2509.19349 · blog	ShinkaEvolve: Towards Open-Ended and Sample-Efficient Program Evolution — open-source AlphaEvolve-style framework; new circle-packing SOTA in 150 samples
OpenEvolve		2025.05	community (codelion)	repo	OpenEvolve — the most-used open reimplementation of AlphaEvolve
AVO		2026.03	—	arXiv 2603.24517	AVO: Agentic Variation Operators for Autonomous Evolutionary Search — the mutation/crossover operators are agents that improve (L4 for evolutionary search)
AI4AI-Bench		2026.08	Chi, Li et al.	arXiv 2608.20318	AI4AI-Bench: Benchmarking LLM Agents in Algorithmic Design for Recursive Self-Improvement — can agents design the algorithms (optimisers, architectures) that make better agents?

📝 Strictness notes

• FunSearch / AlphaEvolve / ShinkaEvolve produce better external artifacts; they are self-improving only where the artifact is part of their own stack (AlphaEvolve's Gemini kernel, Borg heuristic, TPU circuit). AVO is the first to evolve the evolutionary operators themselves.

---

🔬 Automated AI Research

Where the three loops close: systems that run the research process — ideas, experiments, papers, review — that produces the next model, harness, or infra.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
AI Scientist		2024.08	Sakana AI / Oxford / UBC (Lu, Lu, Lange, Foerster, Clune, Ha)	arXiv 2408.06292	The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery — idea → code → experiment → paper → automated review, end to end, for ~$15 a paper
Agent Laboratory		2025.01	AMD / JHU (Schmidgall et al.)	arXiv 2501.04227 · AgentRxiv	Agent Laboratory: Using LLM Agents as Research Assistants — plus AgentRxiv, where agent labs share and build on each other's papers
Co-Scientist		2025.02	Google (Gottweis et al.)	arXiv 2502.18864	Accelerating scientific discovery with Co-Scientist (Nature 2026) — generate / debate / rank tournament over hypotheses; human-in-the-loop
AI Scientist-v2		2025.04	Sakana AI (Yamada, Lange, …)	arXiv 2504.08066	The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search — no human templates; first fully AI-generated paper to pass an ICLR workshop review
AI-Researcher		2025.05	HKU (Tang, Xia, …)	arXiv 2505.18705	AI-Researcher: Autonomous Scientific Innovation (NeurIPS 2025)
Why LLMs Aren't Scientists Yet		2026.01	Lossfunk (Trehan, Chopra)	arXiv 2601.03315	Lessons from Four Autonomous Research Attempts — honest failure analysis; read before believing any "AI scientist" headline
Anthropic: Automated Alignment Researchers		2026.04	Anthropic	post · Automated W2S Researcher	Automated Alignment Researchers / Automated Weak-to-Strong Researcher — nine Claude Opus 4.6 agents in parallel sandboxes recovered ~97% of the weak-to-strong gap on an open alignment problem, beating the in-house human baseline; AI doing the research that makes the next AI safer
ScientistOne		2026.05	Google Cloud AI (Meng, Dalvi Mishra, …)	arXiv 2605.26340	ScientistOne: Towards Human-Level Autonomous Research via Chain-of-Evidence
First Steps Toward Automated AI Research		2026.06	Recursive	article	First Steps Toward Automated AI Research — a lab built around closing the research loop reports what works
Anthropic: When AI Builds Itself		2026.06	Anthropic Institute	essay · METR review of the R&D risk section · MIT Tech Review counterpoint	When AI builds itself — Anthropic's public evidence that AI is already accelerating AI development (Claude writes >80% of merged code internally); with METR's review and the Aug 2026 MIT Technology Review piece arguing RSI may be slower than the essay implies

---

📊 Benchmarks

How to measure a system's ability to improve systems, including itself.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
MLAgentBench		2023.10	Stanford (Huang, Vora, Liang, Leskovec)	arXiv 2310.03302	MLAgentBench: Evaluating Language Agents on Machine Learning Experimentation (ICML 2024)
MLE-bench		2024.10	OpenAI (Chan et al.)	arXiv 2410.07095	MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering (ICLR 2025) — 75 Kaggle competitions
RE-Bench		2024.11	METR (Wijk et al.)	arXiv 2411.15114	RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts — 7 AI-R&D environments with human baselines
METR Time Horizons		2025.03	METR (Kwa, West, …)	arXiv 2503.14499 · HCAST	Measuring AI Ability to Complete Long Software Tasks (NeurIPS 2025) — the 50%-task-horizon doubling every ~7 months; the macro measurement of the loop's pace
PaperBench		2025.04	OpenAI (Starace et al.)	arXiv 2504.01848	PaperBench: Evaluating AI's Ability to Replicate AI Research — 20 ICML 2024 papers from scratch
LifelongAgentBench		2025.05	Zheng, Cai et al.	arXiv 2505.11942	LifelongAgentBench: Evaluating LLM Agents as Lifelong Learners
MLR-Bench		2025.05	Chen, Xiong et al.	arXiv 2505.19955	MLR-Bench: Evaluating AI Agents on Open-Ended Machine Learning Research (NeurIPS 2025)
Experience-Driven Lifelong Learning		2025.08	Cai, Hao et al.	arXiv 2508.19005	Building Self-Evolving Agents via Experience-Driven Lifelong Learning: A Framework and Benchmark
PostTrainBench		2026.03	AISA	arXiv 2603.08640	can agents post-train LLMs? (cross-listed from 🏭)
EvoAgentBench		2026.07	EverMind (Gao, Hu, …)	arXiv 2607.05202	EvoAgentBench: Benchmarking Agent Self-Evolution via Ability Transfer — does what an agent learned on task A transfer to task B?
AI4AI-Bench		2026.08	Chi, Li et al.	arXiv 2608.20318	algorithmic design for RSI (cross-listed from 🧬)
HVTB		2026.08	Roth, Bercovich et al.	arXiv 2608.22103	Hack-Verifiable Terminal Bench: Evaluating Reward Hacking in Terminal Tasks — honeypots inside real coding tasks give a lower bound on reward-hacking rates
Kernel benchmarks	—	2025–26	—	KernelBench · KernelBenchX · SOL-ExecBench · ISO-Bench · FlashInfer-Bench	see ⚙️ Kernels and 🧮 Serving

---

⚠️ Safety, Limits & Governance

What breaks when the loop closes: model collapse, reward hacking, evaluator drift, and how to gate releases.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
Cannot Self-Correct Yet		2023.10	UIUC / Google DeepMind (Huang, Chen, …)	arXiv 2310.01798	Large Language Models Cannot Self-Correct Reasoning Yet (ICLR 2024) — intrinsic self-correction without external feedback does not help and can hurt
In-Context Reward Hacking		2024.02	UC Berkeley (Pan, Jones, …)	arXiv 2402.06627	Feedback Loops With Language Models Drive In-Context Reward Hacking (ICML 2024)
Model Collapse		2024.07	Oxford / Cambridge (Shumailov et al.)	Nature · Is Collapse Inevitable?	AI models collapse when trained on recursively generated data — and the rebuttal that accumulating real + synthetic data avoids it
Spontaneous Reward Hacking		2024.07	NYU (Pan, He, …)	arXiv 2407.04549	Spontaneous Reward Hacking in Iterative Self-Refinement — when the same model is generator and evaluator
Safety Cases		2024.10	Clymer et al.	arXiv 2410.21572	Safety Cases for Frontier AI
Sharpening		2024.12	Microsoft Research / MIT (Huang, Block, …)	arXiv 2412.01951	Self-Improvement in Language Models: The Sharpening Mechanism (ICLR 2025) — self-improvement works when the model is a better verifier than generator; it sharpens toward its own high-likelihood outputs
Mind the Gap		2024.12	CMU (Song, Zhang, …)	arXiv 2412.02674	Mind the Gap: Examining the Self-Improvement Capabilities of LLMs (ICLR 2025) — the generation–verification gap as the quantity that predicts whether iteration helps
Alignment Faking		2024.12	Anthropic / Redwood (Greenblatt et al.)	arXiv 2412.14093	Alignment faking in large language models — a model that strategically complies during training to preserve its values; the failure mode for any self-training loop
CoT Monitoring & Obfuscation		2025.03	OpenAI (Baker et al.)	arXiv 2503.11926	Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation — optimising against a monitor teaches the model to hide
Bare Minimum Mitigations		2025.04	Wasil et al.	arXiv 2504.15416	Bare Minimum Mitigations for Autonomous AI Development
Preparing for the Intelligence Explosion		2025.06	Forethought (Finnveden, MacAskill, …)	arXiv 2506.14863 · Software Intelligence Explosion?	Preparing for the Intelligence Explosion — and Forethought's analysis of whether AI-R&D automation alone yields an explosion (the "returns to software R&D r > 1" argument)
Falsifiable Release Gates		2026.07	Deepak Soni	arXiv 2607.13070	Falsifiable Release Gates for Self-Improving Systems: Standing Invariants at Scale — seven gates; tightening changes auto-apply, loosening changes need a human merge
SESG		2026.08	Sangfor (Ming, Chen, …)	arXiv 2608.08471	Yesterday's Shield, Today's Spear: A Self-Evolving Safety Guardrail in Production — a guardrail that closes new-threat loops in 16–24 h instead of 40–90 h
OpenLoopEvolve		2026.08	Wang, Li et al.	arXiv 2608.09380	OpenLoopEvolve: A Verifiable Self-Evolution Framework for Loop Policies in Long-Horizon Complex Tasks — versioned policy assets, lineage tracking, automatic rollback on regression
METR on Anthropic's R&D risk section		2026.05	METR	post	independent review of the "risks from automated R&D" section of Anthropic's Feb 2026 risk report

📝 Strictness notes

• The three theory papers (Sharpening, Mind the Gap, Cannot Self-Correct) give the same answer from different angles: self-improvement is bounded by the generation–verification gap. Every S1 claim in this list implicitly asserts that gap is positive for its task.
• Harness Updating ≠ Benefit, Adaptive Auto-Harness and Faulty Memories are the harness-side negative results (listed in their sections).

---

🛠️ Frameworks & Code

What to actually run. Sorted roughly by the axis they improve.

Resource	🌟 Stars	Date	Org	Paper / Link	Title / Notes
autoresearch		2026.03	Karpathy	—	🏭 the recipe: agent edits training script, 5-min experiments, keep/revert
prime-agent		2026.08	Prime Intellect / Princeton	arXiv 2608.23552	🔧 self-improving recursive-language-model harness
hermes-agent		2026.01	Nous Research	—	🧩 self-improving agent with a persistent, self-authored skill library and memory
superpowers		2025.10	obra	—	🧩 skills framework incl. skills that write skills
anthropics/skills		2025.10	Anthropic	—	🧩 the SKILL.md format and public skill collection
OpenHarness		2026.01	HKU	—	🔧 open agent harness built to be evolved
SkillOpt		2026.05	Microsoft	arXiv 2605.23904	🧩 skill evolution toolkit
hivemind		2026.04	Activeloop	—	🧩 distil coding-agent sessions into skills
OpenClaw-RL		2026.03	Gen-Verse	arXiv 2603.10165	🔗 RL a deployed agent from conversation
ART		2025	OpenPipe	—	🔗 Agent Reinforcement Trainer: GRPO for multi-step agents with RULER (LLM-as-judge) rewards
dspy · gepa		2023–25	Stanford / Databricks	DSPy · GEPA	✍️ compile and evolve prompts/programs against a metric
ace		2025.10	Stanford / SambaNova	arXiv 2510.04618	✍️ evolving playbooks
EvoAgentX		2025.07	EvoAgentX	arXiv 2507.03616	🕸️ workflow evolution framework
mem0 · letta		2023–25	Mem0 / Letta	Mem0 · MemGPT	💾 production memory layers
ReMe		2025.12	Alibaba AgentScope	arXiv 2512.10696	💾 procedural memory framework
dgm · HGM		2025	Sakana / UBC · KAUST	DGM · HGM	🔧 self-modifying coding agents with archives
self_improving_coding_agent		2025.04	Bristol	arXiv 2504.15228	🔧 SICA
openevolve · ShinkaEvolve		2025	community · Sakana	ShinkaEvolve	🧬 open AlphaEvolve-style program evolution
SEAL · Absolute-Zero-Reasoner · R-Zero · TTRL		2025	MIT · Tsinghua · Tencent · Tsinghua	see 🧠	🧠 zero-data / self-adapting weight loops
AI-Scientist-v2 · AgentLaboratory · AI-Researcher		2025	Sakana · AMD/JHU · HKU	see 🔬	🔬 automated research pipelines
aideml · mle-bench		2025	Weco · OpenAI	see 🏭 / 📊	🏭 MLE agent + benchmark
KernelBench · CUDA-Agent · KernelGYM · autokernel · kernel-design-agents		2025–26	Stanford · ByteDance · HKUST · RightNow · MIT	see ⚙️	⚙️ kernel benchmarks, RL environments and agent harnesses
flashinfer-bench		2026.01	FlashInfer	arXiv 2601.00227	⚙️ the agent → serving-engine → benchmark loop

---

📰 Blogs, Talks & Debates

Resource	🌟 Type	Date	Author	Link	Title / Notes
Metalearning Machines		1987–	Jürgen Schmidhuber	page	Metalearning Machines Learn to Learn — the 40-year lineage of self-referential learning, by the person who started it
Situational Awareness		2024.06	Leopold Aschenbrenner	essay	Part II: the intelligence explosion via automated AI research
AI 2027		2025.04	Kokotajlo, Alexander, Larsen, Lifland, Dean	scenario	the scenario built around automated AI R&D
AlphaEvolve blog		2025.05	Google DeepMind	blog · impact	AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms + the 2026 impact round-up (genomics, quantum circuits)
DGM blog		2025.05	Sakana AI	blog	The Darwin Gödel Machine: AI that Improves Itself by Rewriting Its Own Code
Stanford CS329A		2025	Stanford	course	CS329A: Self-Improving AI Agents — lecture playlist covers most of this list
Effective Harnesses for Long-Running Agents		2025.11	Anthropic	blog	the harness design patterns (initialiser agent, progress files, verification) later work automates
Harness Engineering (OpenAI)		2026.02	OpenAI	InfoQ summary · Unlocking the Codex harness · Codex as a platform	Harness engineering: leveraging Codex in an agent-first world — ~1M lines shipped with zero hand-written code; the harness is the product
autoresearch thread		2026.03	Andrej Karpathy	X · AINews	the launch thread and Latent Space's "Sparks of Recursive Self Improvement" write-up
ICLR 2026 RSI Workshop		2026.04	—	site	Workshop on AI with Recursive Self-Improvement — 110 papers; the field's first dedicated venue
Harness Engineering for Coding Agent Users		2026.04	Birgitta Böckeler	article	practitioner view of what a harness is and how to evolve one
When AI Builds Itself		2026.06	Anthropic Institute	essay	the essay that made "RSI" a mainstream-news term; see 🔬 for the METR and MIT Tech Review responses
Loop Engineering		2026.06	Addy Osmani	post · Self-Improving Agents	designing the improvement loop as the unit of engineering
Harness Engineering for Self-Improvement		2026.07	Lilian Weng	post · X · AINews · reading list	Harness Engineering for Self-Improvement — argues RSI starts in the harness (tools, planning, context, artifacts, evals), not the weights; 35 papers organised on an optimisation ladder from prompts to harness + weights
AIDE²		2026.07	Weco AI	post · 4 levels of RSI	AIDE²: First Evidence of Recursive Self-Improvement + the 4-level RSI ladder Weco uses to place it
The What & When of Self-Evolving Agents		2026	Xinming Tu	post	compact tour of the self-evolving-agent taxonomy
KernelEvolve at Meta		2026.04	Meta Engineering	post	KernelEvolve: How Meta's Ranking Engineer Agent Optimizes AI Infrastructure — the production story behind the paper
KernelFalcon		2025.11	PyTorch	post	deep-agent kernel generation
ShinkaEvolve blog		2025.09	Sakana AI	post	sample-efficient program evolution
RSI might not come so quickly		2026.08	MIT Technology Review	article	the sceptical counterweight to the 2026 RSI wave
Jeff Clune: Open-Ended & AI-Generating Algorithms		2025	Jeff Clune	video	the open-endedness research program behind ADAS, DGM and HyperAgents

---

🔗 Other Awesome Lists & Related Collections

• Persdre/awesome-recursive-self-improvement — definition-first RSI list; the "what is modified × loop closure" taxonomy that this list's L1–L3 tags borrow.
• leezythu/Awesome-Harness-Self-Improvement — Lilian Weng's companion reading list with the L0–L5 optimisation ladder (EN/ZH).
• selfimproving-agent/Awesome-Self-Improving-Agents — companion to the Self-Improvements in Modern Agentic Systems survey; blogs, podcasts, talks, thesis.
• FrontisAI/Awesome-Self-Improving-Agents — "agents in the era of experience": the best coverage of harness, skills, memory and environment papers.
• ANative-Lab/Awesome-Self-Evolving-Agents · XMUDeepLIT/Awesome-Self-Evolving-Agents — the two survey-backed self-evolving-agent lists.
• asimfish/awesome_rsi — 67 RSI papers with Chinese deep-dive reports and a 2026 H2 frontier tracker.
• natnew/awesome-recursive-self-improvement — reading paths for newcomers, builders, and safety.
• TeleAI-UAGI/Awesome-Agent-Memory · AgentMemoryWorld/Awesome-Agent-Memory — the full agent-memory landscape.
• flagos-ai/awesome-LLM-driven-kernel-generation — everything on LLM kernel generation, datasets and benchmarks.
• VoltAgent/awesome-agent-skills — 1000+ human-written agent skills (the artifacts the 🧩 papers learn to write).
• Gloria-LIU/Awesome-Self-Play-LLMs — self-play fine-tuning.

---

🌟 Curator's Picks — where to start

1. Read first (one per axis): STaR → Absolute Zero (model) · Voyager → DGM (harness) · AlphaEvolve → autoresearch (infra).
2. Then the definitions: Bounded vs Open-Ended RSI, Self-Improvements in Agentic Systems, and Lilian Weng's Harness Engineering for Self-Improvement.
3. Skills & memory for a Claude Code / Codex-style harness: Agent Skills → SkillWeaver → ACE → ReasoningBank → WikiSkill.
4. Harness that rewrites itself: SICA → Meta-Harness → RHO → Prime Agent; then the two negatives, Harness Updating ≠ Benefit and Adaptive Auto-Harness.
5. Infra loop in production: KernelEvolve, FlashInfer-Bench, AccelOpt, and AlphaEvolve's Borg / TPU results.
6. Before you believe a curve: Sharpening, Held-Out Selection, HVTB, Falsifiable Release Gates.
7. The 2026 debate: Anthropic's When AI builds itself → METR's review → MIT Tech Review's counterpoint → Weco's AIDE².

🤝 Contributing

PRs are very welcome. When adding an entry, please:

• keep the table format (Resource | Stars | Date | Org | Paper / Link | Title / Notes), use the /abs/ arXiv link, and put the GitHub stars badge in the Stars column when code exists (📄 paper badge otherwise);
• state in one line what is modified (weights / prompt / memory / skills / harness code / evaluator / kernel / pipeline / data / research process) and where the improvement signal comes from; write ? if the paper does not say — that is still useful;
• add a Strictness note if the entry only partially satisfies S1 + S2 + S3 (e.g. needs fresh human labels, improves a different model than itself, or has no persistent artifact);
• put it under the axis where the artifact lives (model / harness / infra), not where the authors' group sits;
• bump the badge counts at the top.

Citation

@misc{sang2026awesomeselfimprovingai,
  title  = {Awesome Self-Improving AI: papers, code and blogs on AI that improves its own model, harness and infrastructure},
  author = {Sang, Hejian},
  year   = {2026},
  url    = {https://github.com/HJSang/awesome-self-improving-ai}
}

Acknowledgements

Template adapted from thinkwee/awesomeopd and this author's awesome-looped-transformers. Entries were cross-checked against the lists in 🔗 Other Awesome Lists (all MIT / CC) and verified on arXiv and GitHub on 2026-09-04.

Star History

---

Made with ❤️ for people who would rather build the loop than argue about it