LEASH: Uncertainty-Aware Weak-to-Strong Escalation and Oversight for Computer-Use Agents
LEASH escalates computer-use agents by latent uncertainty, matching Claude Opus 5 at 19% lower cost.
LEASH is a training-free framework that lets a weak computer-use agent act by default and escalates to a strong model only when latent uncertainty is high. Uncertainty is estimated by comparing the weak model's hidden states with past successful and failed trajectories, reusing activations already computed for action generation. After handoff, a shadow observer continues to monitor the black-box strong agent. On OSWorld, LEASH with Qwen3.5-35B-A3B reached 73.66% task success at $0.4117 per task, matching always-on Claude Opus 5 (72.90%) while cutting inference cost by 19.2%, and beat random escalation by 6.55–11.94 points.
52