Let AI branches explore freely, but make authority vest slowly
Recursive AI systems can create specialist branches, but risk changes when a branch asks to send data or deploy code. Progressive Risk Vesting says exploration is cheap inside the sandbox. Authority is not. It should be granted sparingly, charged clearly, and bounded over the whole path.
What the source reports
The arXiv abstract describes recursive AI systems that broaden search by spawning specialists. These systems can make new branches to test more paths. Some branches later request tools that send data or deploy code. The paper separates sandbox spawning from capability activation. Sandbox branches stay behind outside controls. Activation is different. A chosen branch crosses a boundary where action may not be fully undone. Progressive Risk Vesting holds a risk budget in escrow for the whole path. It debits the budget as branches activate. The paper proves an anytime harm bound for adaptive trees under stated gate, charge, and compute conditions. A stylized branching model introduces an authority reproduction number. The studies are synthetic and do not estimate deployed-AI safety.
Original source
- Title
- Spawn Freely, Act Sparingly: Progressive Risk Vesting for Recursive LLM-Agent Trees
- Publication
- arXiv