RSI-Anything: Unlocking Recursive Self-Improvement with World Models
The paper proposes a self-improvement loop where world models help decide which actions and updates are worth keeping.
RSI-Anything builds on JEPA-Anything to guide action selection, choose learning data, and adapt the world model when observed feedback supports it. The authors report gains in code repair, GUI task completion, lab planning, reaction optimization, and an immune-science workflow. They also describe checks meant to retain only updated policy-model pairs that remain compatible. Source: HF Daily Papers' note.
RSI-Anything builds on JEPA-Anything to guide action selection, choose learning data, and adapt the world model when observed feedback supports it. The authors report gains in code repair, GUI task completion, lab planning, reaction optimization, and an immune-science workflow. They also describe checks meant to retain only updated policy-model pairs that remain compatible. Source: HF Daily Papers' note.
score 5