LEASH: Uncertainty-Aware Weak-to-Strong Escalation and Oversight for Computer-Use Agents
LEASH reports matching an always-on frontier agent’s task success while cutting inference costs by 19.2%.
On OSWorld, LEASH with Qwen3.5-35B-A3B achieved 73.66% task success at $0.4117 per task, compared with 72.90% for always-on Claude Opus 5. The training-free framework routes tasks and escalates individual steps by comparing the weaker model’s hidden states with those from past successes and failures. It reuses action-generation activations without an extra forward pass, then uses a shadow observer to oversee the stronger agent after handoff. Across four model backbones, it beat random escalation by 6.55–11.94 percentage points at matched escalation rates.
ArXiv · AI/CL/LG's note
On OSWorld, LEASH with Qwen3.5-35B-A3B achieved 73.66% task success at $0.4117 per task, compared with 72.90% for always-on Claude Opus 5. The training-free framework routes tasks and escalates individual steps by comparing the weaker model’s hidden states with those from past successes and failures. It reuses action-generation activations without an extra forward pass, then uses a shadow observer to oversee the stronger agent after handoff. Across four model backbones, it beat random escalation by 6.55–11.94 percentage points at matched escalation rates.
ArXiv · AI/CL/LG's note
score 5