# RSI-Anything: one prediction and feedback cycle across five research domains

> RSI-Anything is a Python framework from Gen-Verse that applies world-model predictions to rank candidate actions before external feedback is spent, across code repair, GUI interaction, laboratory planning, reaction optimization and immune science.

**Gen-Verse/RSI-Anything** — RSI-Anything: Unlocking Recursive Self-Improvement with World Models

- Repository: https://github.com/Gen-Verse/RSI-Anything
- Website: https://arxiv.org/abs/2610.13912
- Stars: 821 · Forks: 65
- Language: HTML
- License: Apache-2.0
- Published: 2026-10-20 · Updated: 2026-10-20 · Language: en
- Canonical page: https://hysenlabs.com/projects/gen-verse-rsi-anything

## Five domains, one cycle of propose, predict and evaluate

RSI-Anything is a domain-general framework that turns world-model predictions and interaction feedback into recursive self-improvement, and it is built on JEPA-Anything. The audience is research teams that work across more than one task family and want a single process rather than a separate pipeline for each. The stated problem is that using a model means spending real feedback, whether that is running tests, clicking through a GUI or heating a reaction flask, and the framework tries to spend that budget more selectively. The overview applies the process to code repair, GUI interaction, laboratory planning, reaction optimization and immune science. Each domain supplies its own state, actions, prediction targets and feedback, while sharing the same cycle of proposing candidates, anticipating consequences, evaluating selected actions and retaining improvements for later rounds. The slogan on the page compresses the idea into three verbs: predict outcomes, learn from interaction, improve the next decision.

## Prediction ranks candidates before the feedback budget is spent

The mechanism is a six-stage loop, stated as propose, predict, select, evaluate, learn, retain. A policy or workflow proposes actions. The world model predicts their consequences. Selected actions then receive external feedback, which supports learning and evaluation, and the retained result guides the next round. The point of the predict stage is ordering: cheap predicted outcomes decide which candidates are worth spending a real evaluation on. In the code repair domain, that becomes ranking candidate programs by predicted execution outcomes and refining them with test feedback. In reaction optimization, the same shape ranks catalyst, ligand, base and solvent choices by predicted yield, after which measured yields steer later experiments. Because the framework is prediction-led, the quality of the world model is the ceiling on how much the loop can improve, and the JEPA-Anything base is what carries that burden.

## Domain adapters carry the states, actions and feedback

What makes the framework reusable is the adapter layer rather than the loop itself. The documentation states that domain adapters connect the shared workflow to each task's states, action space and external feedback, so the six stages stay fixed while the environment changes underneath. The domains list names the environments in full: CodeContests, APPS Introductory and LiveCodeBench for code repair; MiniWoB++ and OSWorld for GUI interaction; LabHorizon F200 for laboratory planning; Suzuki-Miyaura reaction screens for reaction optimization; and scientific-agent workflows with intervention analysis for immune science. Each of those directories also has a Hugging Face resource tree under the project dataset. One caveat is worth stating plainly: the adapters are described, not shown, so an engineer planning an adoption has to budget for writing or auditing the adapter for their own environment before the shared cycle means anything.

## The rsi package and its rsi console script

The project is packaged as a standard Python distribution with setuptools, and the packaging metadata is the clearest available entry point. It requires Python 3.10 or newer and declares no runtime dependencies of its own:
```toml
[project]
name = "rsi-anything"
version = "0.2.0"
requires-python = ">=3.10"
dependencies = []
```
The distribution installs a console script named rsi, mapped to rsi_anything.cli:main, so once the package is installed the entry point is invoked as rsi. Packages are discovered under src, and catalog JSON files under rsi_anything/catalogs are shipped as package data, which is where per-domain configuration is likely to live. The quick start section of the project page is not reproduced here, so the exact command sequence for a first run is not confirmed; the repository also carries scripts/, configs/, tests/, tasks/ and docs/ directories, including benchmark settings in docs/benchmarks.md that define task definitions, comparison protocols and metric interpretation. Read those settings before comparing any number below against another system.

## Code repair: four revisions per round and three feedback rounds

The code repair results are the most detailed part of the evaluation, and the protocol is specific enough to check. The policy proposes four revisions per round, with up to three feedback rounds. Online tests provide the repair feedback, while final success is measured on the benchmark's own evaluation tests, and results are averaged over three seeds. On 200 CodeContests tasks with three seeds, the reported result is 31.2% final hidden-test success. On APPS Introductory with 1,000 problems it is 46.37% final full-test success, and on LiveCodeBench v6 with 1,055 problems it is 33.05% final full-test success. The distinction between online test feedback and hidden evaluation tests matters when reading these numbers, because the loop is allowed to see the first and is scored on the second.

## GUI, chemistry and immune results use different measurement shapes

Outside code repair, the metrics are not directly comparable with each other, and each one carries a different caveat. On the MiniWoB++ click-checkboxes task, the reported figure is 99% task success with 39.4% fewer actions than the baseline, so the interesting part is the action reduction rather than the success rate. On OSWorld the number is an 18.4% relative success-rate gain over the initial agent, which is a comparison against a starting point rather than an absolute score. Laboratory planning reports a 46.01% planning score on LabHorizon F200 Level 2. Reaction optimization reports a 96.15% mean best observed yield across 100 Suzuki flow LC-MS runs, with all runs finding the pool optimum by observation 10. The immune science entry is different in kind: it nominates a prospective intervention combining SIRP-alpha inhibition with IL-36-gamma-axis modulation for experimental testing, which is a hypothesis rather than a score.

## Apache-2.0 with a licence status file and a dated asset release

The licence is Apache-2.0, and the repository is unusually explicit about licensing paperwork. Alongside the standard LICENSE and NOTICE files there is a LICENSE_STATUS.md and a THIRD_PARTY_NOTICES.md, and the packaging configuration lists all three of those plus the LICENSE as licence files, so they travel with an installed distribution rather than being repository decoration. That is a signal to check what those files say about bundled JEPA-Anything components and dataset terms before redistributing anything. The repository also ships a SHA256SUMS.json for verification. On the maintenance side it is not archived and the last push was on 2026-10-13, and one release is published, assets-v1, dated 2026-10-10 and described as code repair reproduction assets. The description names a paper on Arxiv as the project homepage, and the primary language is listed as HTML because the documentation is the main content of the repository.

## Conclusion

RSI-Anything makes sense for research groups that want one loop for proposing, predicting, evaluating and retaining across several task families, and for teams willing to write a domain adapter for each environment. It does not suit production use where a single evaluated metric matters, because the strongest results come from benchmark settings and one offline experiment set. Before adopting it, read LICENSE_STATUS.md alongside LICENSE, check the task README files under tasks/ for the exact protocol behind each reported figure, and confirm that the JEPA-Anything prediction base it depends on is available in your environment.

## FAQ

### What does RSI-Anything actually do in a single round?

The framework runs a six-stage cycle of propose, predict, select, evaluate, learn and retain. A policy proposes actions, the world model predicts their consequences, selected actions get external feedback, and the retained result guides the next round.

### Which Python versions and packages does RSI-Anything need?

The packaging metadata in pyproject.toml requires Python 3.10 or newer and lists no runtime dependencies, with the package named rsi-anything at version 0.2.0. Installing it provides a console script called rsi that points at rsi_anything.cli:main.

### How do I check whether the RSI-Anything benchmark numbers are comparable with my own setup?

The project points to docs/benchmarks.md for task definitions, comparison protocols and metric interpretation, and each domain has its own task README under tasks/. The code repair protocol itself proposes four revisions per round with up to three feedback rounds, with online tests used for feedback and hidden evaluation tests used for the final score.

### Is RSI-Anything suitable for production use, and is it licensed permissively?

The licence is Apache-2.0, with additional LICENSE_STATUS.md and THIRD_PARTY_NOTICES.md files shipped with the package. The published results are benchmark and laboratory experiment figures rather than production measurements, so the framework itself is the research component rather than a finished service.

## Sources

- [Gen-Verse/RSI-Anything on GitHub](https://github.com/Gen-Verse/RSI-Anything)
- [License: Apache-2.0](https://github.com/Gen-Verse/RSI-Anything/blob/main/LICENSE)
- [Project website](https://arxiv.org/abs/2610.13912)
- [README](https://github.com/Gen-Verse/RSI-Anything/blob/main/README.md)
- [Releases](https://github.com/Gen-Verse/RSI-Anything/releases)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/gen-verse-rsi-anything