LEASH framework for uncertainty-aware weak-to-strong agent escalation
October 14, 2026
LEASH is a training-free framework that uses an agent's latent uncertainty to trigger handoffs from a small, cheap model to a frontier model. It maintains control over the strong model after handoff to prevent it from losing track of long-horizon tasks.
HOW THIS AFFECTS YOU
●
builderYou can use this approach to reduce inference costs in long-horizon agent tasks without sacrificing accuracy.
●
researcherThis provides a methodology for managing the handoff between weak and strong models using latent uncertainty signals.