MetaRSI / RSI²
A Meta-Recursive Self-Improving System for Recursive Self-Improving Systems Themselves
Recursive self-improvement (RSI) has so far been verified almost only on machine-scorable benchmarks such as code and math. MetaRSI-v1 breaks improvement into three typed operators: scaling up data, rewriting scaffolds, and training parameters. The three share one loop kernel; an optimizer decides the order, and a meta-level policy then rewrites the schedule itself across rounds. There is no external teacher at any point: the target model plays every role in the loop itself.
01Finding: what constrains RSI is the verification format, not the discipline
The paper opens with a census. Among the 45 RSI systems surveyed, 69% close their improvement loops against closed, machine-checkable objectives: test sets, exact-match answers, numerical targets, or multiple-choice labels. Grouped by discipline, almost every science-domain system falls inside this closed block.
Widened to 22 domains and 55 improvement loops: the number of improvement loops a domain can sustain falls as the verifier level of that domain falls, Spearman ρ = −0.75, p < 0.001. Of the 55 improvement loops, 53 sit on levels 1 through 3 of the verifier ladder.
The conclusion is direct: how far self-improvement can go depends on whether a reliable verifier exists, not on which discipline the task belongs to. The paper therefore argues that the next step for RSI is to enter real and varied science, engineering, and meta-science domains, rather than stop on the small, easily scored patch.
02Method: one kernel, three operators, three nested layers
MetaRSI-v1 splits a system into three writable surfaces: data 𝒟, scaffold ℋ, and parameters θ. Each operator writes exactly one surface, so every adopted change can be attributed to a single operator.
| Operators | What it writes | What it does |
|---|---|---|
| Data-RSI | Synthetic data 𝒟 | Amplifies existing capability and marks where it ends |
| Harness-RSI | Five-slot scaffold ℋ | Rewrites the task scaffold without touching weights |
| Model-RSI | Parameters θ | Internalizes capability into parameters through bounded training |
The three operators share one loop kernel and one vocabulary of artifacts, so they compose instead of excluding each other. Five of the six ordered transitions are legal; one dual-axis optimizer decides both the order of operators and the proposal strategy of each operator. Two layers sit above: the RSI² Agent runs one improvement round, and the MetaRSI² Agent rewrites the scheduling policy itself once per round. That is what the title means by improving recursive self-improving systems themselves.
The loop has one clear permission boundary: the model only proposes. Judging, producing, and reading are done by deterministic code.
03Results
The main experiment takes Qwen3.5-35B-A3B as the target model, evaluated on the sealed test sets of four benchmarks (for GPQA, the pre-frozen 100 hard problems GPQA-D-hard100), with an identical budget for all conditions; results are the mean of 5 outer seeds.
- MetaRSI-v1 wins on all four benchmarks: on average +10.9 points over the frozen system, and +3.6 points over the strongest fixed-order combination. Both fixed-order combinations stop at +7.3 while single operators differ by 3.8 points; scheduling alone is worth 3.6 points.
- Six frontier models each run the MetaRSI-v1 loop on Terminal-Bench 2.1, acting as their own proposer and judge; all improve, by +5.6 to +9.2 on average +7.3。
- Scaffolding-only route (no weight updates) over three consecutive rounds: +7.3, +2.8, +1.2 per round on average, cumulatively +11.3。
- Five-round comparison: the meta-update path accumulates +26.2, against +20.7 without, and the gap widens every round. An improver that has been improved really is a better improver.
04Significance
The paper redefines self-improvement from editing one surface to composing along the whole model production pipeline, and lays out two routes: the model route internalizes capability through training; the scaffold route leaves the weights untouched, so it scales to any model reachable only through an interface. Data-RSI is the shared base of both routes. The paper also gives falsifiable regularities: where loops exist, how operators compose, and what supervision buys.
For us, the first finding is the one that matters most: the ceiling of self-improvement is set by the verifier. At the low end of the verifier ladder, the most valuable output of Data-RSI is a boundary statement of where human or corpus supervision is really needed. That is exactly where expert task data sits.
09Citation
@article{tan2026metarsi,
title = {MetaRSI / RSI2: A Meta-Recursive Self-Improving System for Recursive Self-Improving Systems Themselves},
author = {Tan, Zihan and Sun, Leixin and Shi, Zitong and Liu, Yitao and Wu, Jiajun and others},
journal = {arXiv preprint arXiv:2609.06396},
year = {2026}
}