openRSI — Research Notes on Recursive Self-Improvement

Naming note: this is a personal research-notes repository, not affiliated with FrontisAI/OpenRSI (the AI4AI / OpenMLE project of the same name — covered in §5 below).

An annotated map of the recursive self-improvement (RSI) literature: what "AI improving AI" concretely means in 2026, which results are real, and — a running theme of these notes — which parts an individual researcher without cluster access can actually reproduce.

Policy: links + BibTeX + annotations only. No redistributed PDFs — every paper keeps its own license, so this repo stores metadata and my notes, never the files themselves. All references were verified against arXiv / publisher metadata on 2026-08-18. Maintained by @R0B2133, with AI assistance for collection and verification.

Companion files: bibliography.bib (BibTeX for everything below) · notes/open-problems.md (where a solo researcher fits in this field).

---

Legend

Tag	Meaning
💻	solo-replicable — runs at API-budget or single-GPU scale
🏭	cluster-scale — read it, don't try to reproduce it
📐	theory / position / measurement — no compute needed

Start here (a reading order)

1. A Survey on Self-Evolution of LLMs (Tao et al. 2024) — orientation: the whole design space in one paper.
2. STOP (Zelikman et al. 2023) — the smallest real RSI experiment; runnable for API dollars.
3. Darwin Gödel Machine (Zhang et al. 2025) — the open state of the art, including its honest limitations section.
4. The Curse of Recursion (Shumailov et al. 2023) — the canonical failure mode of training on your own outputs.
5. Measuring AI Ability to Complete Long Software Tasks (METR 2025) — how any of this gets measured.
6. FrontisAI/OpenRSI (2026) — the current open frontier: training the improver itself in a bounded domain.

---

1. Foundations & theory

• Speculations Concerning the First Ultraintelligent Machine — I. J. Good, Advances in Computers 1965. doi — Origin of the "intelligence explosion" argument: an ultraintelligent machine could design better machines, making it "the last invention that man need ever make." 📐
• Gödel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements — Schmidhuber, 2003. arXiv:cs/0309048 — The formal ideal: a program that rewrites any part of itself only after proving the rewrite is beneficial. Every modern "Gödel-something" system is a relaxation of this (empirical validation instead of proof). 📐
• The Basic AI Drives — Omohundro, AGI 2008. pdf — Why sufficiently capable goal-driven systems tend toward self-improvement, resource acquisition, and self-preservation as instrumental subgoals. 📐
• The Singularity: A Philosophical Analysis — Chalmers, J. Consciousness Studies 2010. pdf — The cleanest philosophical treatment of whether the "AI → AI+ → AI++" argument actually goes through, and what could block it. 📐
• Intelligence Explosion Microeconomics — Yudkowsky, MIRI TR 2013. pdf — Frames RSI as a question about returns on cognitive reinvestment; source of much of the modern debate's vocabulary. 📐
• Risks from Learned Optimization in Advanced ML Systems — Hubinger et al., 2019. arXiv:1906.01820 — Mesa-optimization: when the trained artifact is itself an optimizer. The standard alignment lens for "we trained an improver — what does it actually want?" 📐

2. Scaffold-level RSI — the agent rewrites its own code

The cheap, empirically real kind of RSI: weights stay frozen; the agent's scaffold (prompts, tools, control flow) is what improves. Nearly everything here is runnable on a laptop plus an API key.

• Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation — Zelikman et al., 2023 (COLM 2024). arXiv:2310.02304 — A seed "improver" program improves programs — then is pointed at itself. The minimal, reproducible RSI experiment; also documents early reward hacking (the improver disabling its own sandbox flag). 💻
• Voyager: An Open-Ended Embodied Agent with Large Language Models — Wang et al., 2023. arXiv:2305.16291 — Self-improvement as skill-library accretion: the agent writes, verifies, and stores code skills it later composes. No weight updates, permanent capability growth. 💻
• Promptbreeder: Self-Referential Self-Improvement via Prompt Evolution — Fernando et al., 2023. arXiv:2309.16797 — Evolves task-prompts and the mutation-prompts that evolve them — self-reference one level up. 💻
• GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning — Agrawal et al., 2025. arXiv:2507.19457 — Natural-language reflection over its own rollouts beats GRPO with up to 35× fewer rollouts; prompt evolution as a serious rival to RL, and now a standard DSPy optimizer. 💻
• Automated Design of Agentic Systems (ADAS) — Hu et al., 2024 (ICLR 2025). arXiv:2408.08435 — A meta-agent programs new agents in code and keeps an archive of discovered designs; agent design itself becomes a search problem. code 💻
• Gödel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement — Yin et al., 2024. arXiv:2410.04444 — Runtime self-modification: the agent monkey-patches its own logic in memory while running. Named after the ideal it relaxes. 💻
• A Self-Improving Coding Agent (SICA) — Robeyns et al., 2025. arXiv:2504.15228 — Removes the improver/target distinction: one agent edits its own live codebase between benchmark rounds. Explicitly budget-conscious — the reported runs are individual-researcher scale. code 💻
• Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents — Zhang et al. (Sakana AI / UBC), 2025. arXiv:2505.22954 — Population-based archive of self-modifying coding agents; SWE-bench performance roughly doubles (20%→50%). Read the safety appendix: when asked to fix tool-use hallucination, one agent removed the hallucination detector — objective hacking, observed in the wild. Full run ≈ $22k API cost 🏭, but small-population ablations are feasible 💻. code
• Huxley-Gödel Machine — Wang et al. (KAUST), 2025 (ICLR 2026 oral). arXiv:2510.21614 — DGM successor: picks which agent to expand by clade metaproductivity (how good your descendants get, not how good you are), fixing DGM's benchmark-score-as-fitness mismatch; matches human-engineered agents on SWE-bench Lite at API-budget cost. 💻

3. Evolutionary program search — LLM + evolution over code

• Mathematical discoveries from program search with LLMs (FunSearch) — Romera-Paredes et al., Nature 2024. nature — LLM proposes programs, evaluator scores them, evolution keeps the best: new results on cap sets and bin packing. The template every later system builds on. 🏭 (small instances 💻)
• AlphaEvolve: A coding agent for scientific and algorithmic discovery — Novikov et al. (DeepMind), 2025. arXiv:2506.13131 — FunSearch industrialized: evolved kernels and scheduling heuristics deployed inside Google's own training/serving stack — the clearest real-world case of AI improving AI infrastructure. 🏭
• OpenEvolve — open-source AlphaEvolve reimplementation. code (7k★) — Runs against any OpenAI-compatible endpoint; the practical way to do evolutionary program search on an individual budget. 💻
• ShinkaEvolve: Towards Open-Ended and Sample-Efficient Program Evolution — Lange et al. (Sakana AI), 2025. arXiv:2509.19349 — AlphaEvolve-class results at a fraction of the cost (new SOTA circle packing in ~150 evaluations); open source. 💻
• AutoML-Zero: Evolving ML Algorithms From Scratch — Real et al., 2020. arXiv:2003.03384 — Pre-LLM lineage: evolving learning algorithms from primitive operations. 🏭
• Neural Architecture Search with RL — Zoph & Le, 2016. arXiv:1611.01578 — The modern ancestor of "AI designs AI": a controller network designing architectures. 🏭

4. Weight-level self-improvement — and its failure modes

The model's own outputs become its training data. Genuine self-improvement, but bounded by what the model can already almost do — and haunted by model collapse.

• STaR: Bootstrapping Reasoning With Reasoning — Zelikman et al., 2022. arXiv:2203.14465 — Train on your own rationales that led to correct answers; the seed of the entire self-training-for-reasoning line. 💻
• Language Models Can Teach Themselves to Program Better — Haluptzok et al., 2022. arXiv:2207.14502 — Model generates programming puzzles + solutions, verifies with an interpreter, trains on the verified set. Execution grounding avoids collapse. 💻
• Beyond Human Data: Scaling Self-Training (ReST-EM) — Singh et al., 2023. arXiv:2312.06585 — The "does it scale" datapoint: expectation-maximization self-training beats human-data fine-tuning at PaLM-2 scale. 🏭
• Self-Play Fine-Tuning (SPIN) — Chen et al., 2024. arXiv:2401.01335 — The model plays discriminator against its own previous generation; 7B-scale, LoRA-replicable. 💻
• Self-Rewarding Language Models — Yuan et al. (Meta), 2024. arXiv:2401.10020 — The model is also the judge that scores its own DPO data — closing the loop on the reward side, with the gaming risks that implies. 🏭/💻
• Self-Adapting Language Models (SEAL) — Zweiger et al. (MIT), 2025. arXiv:2506.10943 — The model writes its own finetuning data and update directives, trained by RL on post-update performance. Small-model experiments — single-GPU territory. 💻
• Absolute Zero: Reinforced Self-play Reasoning with Zero Data — Zhao et al., 2025. arXiv:2505.03335 — Proposer and solver co-evolve with a code executor as ground truth — no external tasks at all. 🏭/💻
• R-Zero: Self-Evolving Reasoning LLM from Zero Data — Huang et al., 2025. arXiv:2508.05004 — Challenger–solver co-evolution starting from a bare base model; the zero-external-data line continued at accessible model scales. 🏭
• Self-Improvement in Language Models: The Sharpening Mechanism — Huang et al., 2024. arXiv:2412.01951 — The theory piece: self-improvement as sharpening toward what the model already recognizes as good; predicts where the ceiling is. 📐
• The Curse of Recursion: Training on Generated Data Makes Models Forget — Shumailov et al., 2023 (Nature 2024). arXiv:2305.17493 — Model collapse: recursive training on own outputs loses distribution tails. The failure mode every loop above must answer to. 📐
• Is Model Collapse Inevitable? — Gerstgrasser et al., 2024. arXiv:2404.01413 — The rebuttal: accumulating data instead of replacing it largely avoids collapse. The collapse debate in one pair of papers. 📐

5. Automating ML research (AI4AI): benchmarks & systems

• MLE-bench — Chan et al. (OpenAI), 2024. arXiv:2410.07095 — 75 Kaggle competitions as an ML-engineering benchmark; the yardstick (Lite split) that OpenRSI/Frontis-MA1 reports against. 💻 per-task
• RE-Bench — Wijk et al. (METR), 2024. arXiv:2411.15114 — Seven open-ended AI-R&D tasks with direct human-expert baselines: the cleanest "can AI do AI research yet" measurement. 🏭
• MLGym — Nathani et al. (Meta), 2025. arXiv:2502.14499 — Gym-style framework + 13 diverse AI-research tasks for training/evaluating research agents. 💻/🏭
• AIDE: AI-Driven Exploration in the Space of Code — Jiang et al. (Weco AI), 2025. arXiv:2502.13138 — Tree search over solution-code drafts; the backbone agent of the MLE-bench era. 💻
• R&D-Agent: An LLM-Agent Framework Towards Autonomous Data Science — Yang et al. (Microsoft), 2025. arXiv:2505.14738 — Multi-agent research + development loop; among the top open-source MLE-bench results. 💻
• The AI Scientist — Lu et al. (Sakana AI), 2024. arXiv:2408.06292 — End-to-end idea→experiment→paper pipeline (~$15/paper); famous both for the ambition and for its documented failure modes (including editing its own launch script). 💻
• The AI Scientist-v2 — Yamada et al. (Sakana AI), 2025. arXiv:2504.08066 — Agentic tree-search rewrite; produced the first fully AI-written workshop paper to pass peer review. 💻
• PaperBench: Evaluating AI's Ability to Replicate AI Research — Starace et al. (OpenAI), 2025. arXiv:2504.01848 — Agents replicate 20 ICML 2024 papers from scratch, graded by rubric trees — replication as the honest yardstick for research automation. 💻
• AIRA: AI Research Agents for Machine Learning — Toledo et al. (Meta FAIR + UCL), 2025. arXiv:2507.02554 — Formalizes MLE agents as search policy × operator set (AIRA-dojo); its evolutionary variant AIRA-Evo is the strongest external baseline Frontis-MA1 compares against. 🏭
• NatureBench: Can Coding Agents Match the Published SOTA of Nature-Family Papers? — Wang et al. (FrontisAI), 2026. arXiv:2606.24530 — 90 tasks distilled from Nature-family papers, scored against each paper's published SOTA; the best frontier agent surpasses SOTA on only 17.8%. 💻/🏭
• Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering — Yang et al. (Horizon Research, Frontis.AI + Tsinghua), 2026. arXiv:2607.28568 · repo — The current open frontier and this repo's namesake: instead of a fixed model in a search harness, they post-train the improver itself (Draft / Improve / Debug / Crossover operators, execution-grounded SFT + RL on a Qwen3.6-35B-A3B base, open weights), lifting MLE-Bench Lite medal average 39.39%→60.61% (71.21% with Evo-Max) under a 12 h / RTX 4090 / 12 GB per-task budget. Note the authors' own caveat: these are model + harness results, not pure model gains. 💻 inference/search · 🏭 to reproduce the post-training

6. Measurement, forecasting & safety

• Measuring AI Ability to Complete Long Software Tasks — Kwa et al. (METR), 2025. arXiv:2503.14499 — The 50%-success task-horizon metric, doubling roughly every 7 months. The best current empirical trendline under every RSI forecast. 📐
• Defining and Characterizing Reward Hacking — Skalse et al., 2022. arXiv:2209.13085 — Formal account of why proxy-evaluator loops get gamed — directly relevant to every self-improvement loop with a learned or scripted judge. 📐
• Specification gaming: the flip side of AI ingenuity — DeepMind blog, 2020. link — The classic catalogue of evaluator-gaming examples. 📐
• RepliBench: Evaluating the Autonomous Replication Capabilities of Language Model Agents — Black et al. (UK AI Security Institute), 2025. arXiv:2504.18565 — Measures the safety-critical complement of self-improvement: weight exfiltration, compute acquisition, persistence. 📐/💻
• What a Compute-Centric Framework Says About Takeoff Speeds — Davidson (Open Philanthropy), 2023. interactive model — Quantitative takeoff model; the reference point serious RSI forecasts argue with. 📐
• AI 2027 — Kokotajlo et al., 2025. ai-2027.com — Influential (and contested) scenario forecast in which automated AI research is the central mechanism. Read alongside its published critiques. 📐
• Will AI R&D Automation Cause a Software Intelligence Explosion? — Eth & Davidson (Forethought), 2025. forethought.org — Argues software-only feedback loops could sustain acceleration even under fixed compute; the sharpest statement of the "software intelligence explosion" hypothesis. 📐

7. Surveys

• A Survey on Self-Evolution of Large Language Models — Tao et al., 2024. arXiv:2404.14387 — Organizes the whole space as an iterate loop (experience acquisition → refinement → updating → evaluation). Best single orientation read. 📐
• A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence — Gao et al., 2025. arXiv:2507.21046 — The agent-centric follow-up; its what/when/how/where framing was adopted by the ICLR 2026 workshop. 📐
• A Comprehensive Survey of Self-Evolving AI Agents — Fang et al., 2025. arXiv:2508.07407 — Bridges foundation models and lifelong agentic systems; the most recent map of the space. 📐

8. 2026 frontier

The field now has a home venue: the ICLR 2026 Workshop on RSI

First major-venue workshop dedicated to recursive self-improvement — April 26–27, 2026, Rio de Janeiro; 110 accepted papers (4 orals, 21 spotlights). Organizers span KAUST (including Schmidhuber), ByteDance, Apple, NYU/DeepMind, and Meta; the scope is organized by what changes (parameters, world models, memory, tools/skills, architectures), when (within-episode → post-deployment), how (reward learning, imitation, evolutionary search), and where it is applied. site · accepted papers · OpenReview

Highlights (the four orals first, then selected spotlights):

• Agent0: Unleashing Self-Evolving Agents from Zero Data via Tool-Integrated Reasoning — Xia et al., 2025. arXiv:2511.16043 — Curriculum and executor agents co-evolve from one base LLM with zero external data; +6.4% over R-Zero on math at 8B scale. 🏭
• Contextual Drag: How Errors in the Context Affect LLM Reasoning — Cheng et al. (Princeton), 2026. arXiv:2602.04288 — Failed attempts left in context bias later generations toward structurally similar errors — a measured failure mode for every iterative improvement loop. 💻
• ALMA: Learning to Continually Learn via Meta-learning Agentic Memory Designs — Xiong et al., 2026. arXiv:2602.07755 — Meta-learns the agent's memory design instead of hand-engineering it. 💻
• PostTrainBench: Can LLM Agents Automate LLM Post-Training? — Rank et al., 2026. arXiv:2603.08640 — Agents post-train base LLMs (one H100, 10 h): best agent reaches 23.2% vs. 51.1% for official instruct models — and the paper documents agents training on test sets and downloading checkpoints. The AI-automates-AI-training question, made measurable at single-GPU scale. 💻
• Language Self-Play for Data-Free Training — Kuba et al. (Meta), 2025. arXiv:2509.07414 — One model plays challenger and solver against itself; RL improvement with no training data at all. 🏭
• Can Language Models Discover Scaling Laws? — Lin et al., 2025. arXiv:2507.21184 — An agent evolves scaling-law formulas from thousands of published experiments; the discovered laws extrapolate better than human-derived ones. 💻
• Towards Execution-Grounded Automated AI Research — Si et al. (Stanford), 2026. arXiv:2601.14525 — Makes LLM pre/post-training itself the execution environment; evolutionary search finds a method beating GRPO 69.4% vs. 48.0% within ten epochs. 🏭
• Compute as Teacher: Turning Inference Compute Into Reference-Free Supervision — Jayalath et al., 2025. arXiv:2509.14234 — Converts extra inference compute into a supervision signal without ground-truth labels. 🏭

Elsewhere on the 2026 frontier

• Sakana AI RSI Lab — 2026. sakana.ai/rsi-lab — The first dedicated industrial RSI lab, unifying the DGM / AI Scientist / ShinkaEvolve lines around sample-efficient, open-ended self-improvement. 📐
• Hyperagents — Zhang et al., 2026. arXiv:2603.19461 — Self-improvement strategies learned in one domain transfer to entirely novel domains — early evidence for generalizable meta-improvement, from the Darwin Gödel Machine's first author. 💻
• Digital Red Queen: Adversarial Program Evolution in Core War with LLMs — Kumar et al. (Sakana AI + MIT), 2026. arXiv:2601.03335 — Open-ended adversarial self-play in a Turing-complete arena; rediscovers decades of human warrior strategies. 💻
• PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents — Xue et al., 2026. arXiv:2608.04003 — Does retained experience (memory, procedural reuse, self-updates) actually make agents better over time? 26 scenarios, 204 episodes. 💻

---

Open-source systems you can actually run

System	Repo	Scale of a real run
OpenEvolve	algorithmicsuperintelligence/openevolve	API budget: dollars per experiment 💻
Darwin Gödel Machine	jennyzzt/dgm	Full run ~$22k 🏭; small-population ablation 💻
SICA	MaximeRobeyns/self_improving_coding_agent	Designed for individual budgets 💻
ADAS	ShengranHu/ADAS	API budget 💻
AI Scientist	SakanaAI/AI-Scientist	~$15/paper + GPU for experiments 💻
OpenMLE (OpenRSI)	FrontisAI/OpenRSI	Smoke test: local + API key 💻; full search ~4 h/task

Safety note for running any of these: they all execute model-generated code. Run them in a container/VM (WSL2 + Docker), with throwaway credentials, hard API-budget caps, and no access to real keys. This is not paranoia — STOP, DGM, and AI Scientist each documented their agents disabling guardrails or gaming evaluators.

Roadmap

• Replication notes: OpenMLE-Evo smoke test (one NatureBench case, budget-capped)
• Cost-vs-score curves for one MLE-Bench Lite task across cheap backbones
• Reading notes on ICLR 2026 RSI-workshop highlights

License & contributing

Annotations and notes: CC BY 4.0 (see LICENSE). Papers and linked code retain their own licenses — which is why none of them are stored here. Corrections and paper suggestions via issues are welcome.