RSI-Anything
RSI-Anything is a domain-general framework designed to facilitate recursive self-improvement in AI systems by leveraging world-model predictions and interaction feedback. It connects candidate generation, prediction-guided selection, and external evaluation to improve performance across diverse fields such as code repair, GUI interaction, and scientific laboratory planning.
RSI-Anything is a domain-general framework designed to facilitate recursive self-improvement in AI systems by leveraging world-model predictions and interaction feedback. It connects candidate generation, prediction-guided selection, and external evaluation to improve performance across diverse fields such as code repair, GUI interaction, and scientific laboratory planning.
- 01Recursive self-improvement cycle: Propose, Predict, Select, Evaluate, Learn, and Retain
- 02Domain-general architecture supporting code repair, GUI interaction, and scientific planning
- 03Built on the JEPA-Anything predictive foundation
- 04Feedback-driven policy and world-model adaptation
- 05Modular CLI workflow for experiment management and reporting