Gen-Verse's RSI-Anything Keeps AI Agents and World Models in Sync Across 5 Domains
Gen-Verse's open-source framework pairs world-model predictions with real feedback, pushing a single recursive loop across code, GUI, chemistry and biology.
Takeaways
- Gen-Verse released RSI-Anything, an Apache-2.0 framework for recursive self-improvement built on JEPA-Anything world models.
- Single Propose-Predict-Select-Evaluate-Learn-Retain loop runs across code repair, GUI agents, lab planning, chemistry, and immune science.
- Code repair rises from 26.3% to 31.2% with coder frozen; GUI agents cut actions by 39.4%.
- Reaction optimization hits 96.15% mean best yield versus 82.16% for Bayesian optimization at the same budget.
- Key design choice: policy and world model are retained only if compatibility and component checks pass together.
- Paper on arXiv, datasets and checkpoints on Hugging Face.
Gen-Verse has released RSI-Anything, an Apache-2.0 framework that applies one improvement loop to code repair, browser control, laboratory planning, chemistry optimization, and immune-science hypothesis generation. The repository has 5 stars. Its practical contribution is a method for improving agents under limited evaluation budgets while checking that each update remains compatible with the rest of the system.
Within the framework, a policy is the model, strategy, or workflow that proposes actions. A world model predicts or scores the likely consequences of those actions. RSI-Anything treats them as a persistent pair because policy updates change which actions the predictor sees, while predictor updates change which actions the system selects.
Six operations, one retained pair
Each supported domain maps its task-specific machinery onto the same six-stage cycle:
- Propose: The policy or workflow generates candidate actions.
- Predict: The world model estimates each candidate’s outcome.
- Select: The system chooses candidates for external evaluation.
- Evaluate: Tests, browser outcomes, simulations, or experiments provide feedback.
- Learn: The policy, workflow, or world model adapts from that feedback.
- Retain: The updated pair advances only after component and compatibility checks.
The framework builds its predictor on JEPA-Anything, a Joint Embedding Predictive Architecture that estimates outcomes in a learned latent space, meaning a compressed representation learned from data. The predictor can rank candidates without generating complete trajectories, reserving expensive external feedback for actions expected to perform well.