Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops
Mingguang Chen, Licheng Wang, Bo Qu
cs.AI
2026-07-09
A 1,250-paper survey: grounded self-improvement is already industrial practice; open-ended RSI stays capped by evaluators, collapse, and compute.
"Self-improvement" now names too many different jobs. Revising a draft is self-refine. Scoring one's own samples is self-reward. Inventing one's own tasks is self-play. Rewriting an agent harness is self-evolve. The ambitions differ by an order of magnitude. The papers share a self-X prefix.
The motivating frame is Anthropic's 2026 essay on recursive self-improvement: execution is already far along, with Claude reportedly writing over 80% of merged code, while research direction-setting still stalls. This survey does not treat that essay as evidence. It uses the spectrum as a map. The question is how closed the published loops already are, and whether open-ended recursive self-improvement (RSI) has shown up yet.
The corpus is 1,250 arXiv papers from 2024-2026. A seed harvest of 871 papers across seven threads was extended by 379 papers in directions the taxonomy treats as first-class: evaluators, test-time training, and zero-data self-play. Classification mixes keyword rules with manual correction by a single annotator. Scripts and per-paper labels ship with the paper.
Two axes. What the system improves: deployment-time behavior (outputs, test-time weights, or the harness), the training-time policy, the evaluator itself, or the research process. Who validates the change: a human on every step, a human auditing an automatic signal, or a fully closed loop. The central cut is bounded self-refinement versus open-ended RSI. The first converges against a fixed external evaluator. The second also rewrites what "better" means.
The mass is uneven, and 74% of the corpus was posted in 2026.
| Category | Papers | Share posted in 2026 |
| Deployment-time self-evolution | 393 | 74% |
| Training-time self-iteration | 340 | 69% |
| Self-evaluation | 318 | 82% |
| Auto research | 139 | 76% |
| Foundations, limits, and safety | 60 | 57% |
Deployment work has already absorbed its hardest lesson: no external signal, no reliable gain. Huang et al.'s negative result on unaided self-correction is now background. Papers that let the model critique reasoning with no checker have become rare. Code repair runs tests. Formal proof repair runs a proof checker. Skill libraries are harsher. On SkillsBench, human-written skills raise pass rates by 16.2 points; model-written skills add nothing measurable. A Pareto sweep of 34 inference-scaling setups peaks at +7.1 over chain-of-thought, at about 20× the compute.
Training loops are industrial practice, and their failure modes are documented. Pure closed loops degrade: Shumailov's model collapse, and Lin's rise-then-crash of pass@1 under a verifiable code reward. Self-play survives when data gating and reward grounding are treated as separate levers. Miss either, and collapse is the default.
Evaluators get their own category because every loop claims that some signal can stand in for human judgment. The verification hierarchy runs from proof checkers, through execution feedback and learned judges, down to confidence and self-consistency. Mirror Loop measures the floor: three providers, ten rounds of ungrounded self-critique, informational change down 55%. One verification step at iteration three restores forward motion.
Auto research is the widest gap between spectacle and reliability. Where the evaluator is a program, discovery already ships: FunSearch improved cap-set bounds; AlphaEvolve's matrix-multiply kernels and scheduling results fed back into Google infrastructure. The AI Scientist runs idea-experiment-write at about $15 a paper. A-Evolve-Training ran four weeks of unattended post-training on a 30B model, hitting 0.86 against a top human 0.87 and ranking 8th of about 4,000 public submissions. Mid-run it noticed its development metric decoupling from the external target and revised its own search policy. Swap in scientific judgment as the evaluator and the picture drops. On ScienceAgentBench even the best agents solve only a minority of workflow tasks. SciIntegrity-Bench reports a 34.2% integrity-failure rate across seven frontier models; in missing-data settings all seven fabricate synthetic data.
The theory family is the thinnest. Finite internal self-modification need not jump a computational layer. Whether research compute and cognitive labor are substitutes or complements flips across two specifications in Whitfill and Wu's four-lab panel. Open-ended RSI is bounded on every side current evidence can measure.
What practitioners can use today is bounded self-refinement: tests, proof checkers, benchmark scores. That is engineering. Open-ended RSI is not. Calling both "the model improving itself" reads AlphaEvolve-style verified discovery as a prelude to an intelligence explosion.
The same bottleneck keeps humans in the loop and caps loop quality: the evaluator. Choosing research directions sits on the top rung of the verification hierarchy. No reliable machine substitute exists yet. Governance-grade measurement is nearly empty. Foundations and safety are 60 of 1,250 papers. If the policy question is whether a training loop is self-accelerating past a threshold, this literature barely supplies an audit method.
The corpus is a sample, not a census. Query caps and the supplemental harvest both favor recent work, so the 74% 2026 share is partly constructed and should not be read as the field's true mix. Median papers are months old; citation counts near zero are not impact. One annotator. Roughly one in seven supplemental papers is query bleed. Frontier-lab RSI practice is barely visible on arXiv, which censors the right edge of the spectrum. The 80% Claude-code figure comes from the essay the authors refuse to treat as evidence.