RSI AtlasRecursive self improvement, mapped and runnable
Research

Research Map Across Papers

Anchor works and recent advances that shape the atlas, organized from classic theory to current loops.

14 min readAdvancedUpdated 2026-09-18

How to read this map

The field moves fast and uses many near synonyms. This map groups work by what improves and how the loop closes. Behavior loops refine answers at run time. Training loops turn traces into weights. Evaluator work decides what good means. Research loops run experiments. Foundations and safety hold the rest honest.

A July 2026 survey organizes about 1250 arXiv papers from 2024 through 2026 along the same lines. Its central cut guides this atlas. Bounded refinement improves behavior against a fixed external check. Open ended recursion also changes the method and the criteria. The second path is far less proven.

Classic roots

Godel Machines set the classic ideal of self rewrite after proof of utility with global optimality framing. The proof demand is too strict for daily practice, so modern loops use empirical gates with sandboxes and full suites.

STaR showed a simple training loop of generate, filter on correct answers, finetune, and repeat. Self Refine showed one model acting as writer, critic, and reviser. Reflexion added verbal reflection with episodic memory. FunSearch paired code generation with automatic evaluation in evolution. Each pattern reappears in current systems.

Current loop designs

Self Rewarding Models let the same model generate answers and judge them for preference training, which creates a genuinely recursive signal with saturation risk. Godel Agent made runtime self patch concrete with memory read and write plus recursive control.

Darwin Godel Machine added an archive of agents with empirical validation on coding benchmarks and reported strong SWE bench movement. AlphaEvolve extended evolution to full files with multi objective search. MetaEvolve moved repair skill into weights with execution rewards.

Recent RSI papers refine the same ideas. Dream RSI uses replay from discovery trees for cheap off policy feedback. Meta N fixes the meta operator and recurses on growing input with convergence set depth. MetaRSI composes data, harness, and model operators under one kernel. Generalized Agent Iteration gives dials for grounding and self reference. Harness self improvement shows context management gains with lower cost. RSIAgent shows training free memory growth with broad then deep search. The Last AI Built by Humans frames the road from execution autonomy to meta improvement. Karpathy autoresearch gives the minimal public template of one file, one metric, and one time box.

Anchor papers and recent advances

PaperYearCore idea
Godel Machines by Schmidhuber2003Self rewrite after proof of utility
STaR Self Taught Reasoner2022Bootstrap rationales that lead to right answers
Self Refine2023Writer critic reviser loop with one model
Reflexion2023Verbal reflection stored in episodic memory
FunSearch2023Code evolution guarded by automatic evaluator
Self Rewarding Language Models2024Same model generates and judges for preference data
Godel Agent2024Runtime self read and self patch with recursion
Darwin Godel Machine2025Archive search with benchmark validation
AlphaEvolve2025Full file evolution with rich context
MetaEvolve2026RL for multi round repair meta skills
RSI Survey across 1250 papers2026Taxonomy of behavior, policy, evaluator, research
Dream RSI2026Replay simulator for cheap exploration feedback
Meta N2026Fixed meta operator with growing input depth
MetaRSI RSI22026Composed data harness model operators
Generalized Agent Iteration2026Formal frame with grounding dials
Recursive Harness Self Improvement2026Prompt level harness edits from pairwise history
RSIAgent2026Training free memory with broad then deep search
The Last AI Built by Humans2026Roadmap across science, embodied work, software
Karpathy autoresearch2026Minimal keep or discard research loop
Anthropic progress note on RSI2026Industry view of harness model co evolution

Continue reading

Godel Machines classic paper
https://arxiv.org/abs/cs/0309048
STaR Self Taught Reasoner
https://arxiv.org/abs/2203.14465
Self Refine
https://arxiv.org/abs/2303.17651
Reflexion
https://arxiv.org/abs/2303.11366
FunSearch
https://arxiv.org/abs/2305.20050
Self Rewarding Language Models
https://arxiv.org/abs/2401.10020
Godel Agent
https://arxiv.org/abs/2410.04444
Darwin Godel Machine
https://arxiv.org/abs/2505.22954
AlphaEvolve
https://arxiv.org/abs/2506.13131
MetaEvolve
https://arxiv.org/abs/2607.21971
RSI Survey across 1250 papers
https://arxiv.org/abs/2607.07663
Dream RSI
https://arxiv.org/abs/2609.14858
Meta N emergent depth
https://arxiv.org/abs/2608.24735
MetaRSI RSI2
https://arxiv.org/abs/2609.06396
Generalized Agent Iteration
https://arxiv.org/abs/2609.13406
Recursive Harness Self Improvement
https://arxiv.org/abs/2607.15524
RSIAgent
https://arxiv.org/abs/2609.15364
The Last AI Built by Humans
https://arxiv.org/abs/2609.11873
Karpathy autoresearch
https://github.com/karpathy/autoresearch
Code review with spam filter
/docs/code-review
Core loop deep dive
/docs/loop
Found a gap in this page.Open the GitHub repo to propose a fix