LEASH: latent uncertainty enables efficient weak-to-strong escalation for computer-use agents

Reusing internal activations to detect uncertainty without extra forward passes lets lightweight agents escalate to frontier models only when needed and monitor them afterward, cutting OSWorld inference costs by 19%.

Primary source: arXiv:2610.22280: https://arxiv.org/abs/2610.22280

What’s new: Autonomous computer-use agents tackling complex multi-step desktop workflows face an acute efficiency trade-off: executing every action with a frontier reasoning model is prohibitively expensive, whereas relying on compact open-weight models leads to frequent catastrophic errors. Weak-to-strong escalation frameworks offer a compelling compromise by letting a small agent execute routine tasks and handing off control to a strong model only when necessary. However, deciding when to escalate typically requires expensive external verifiers, and once control is transferred, black-box commercial APIs offer no internal telemetry for runtime safety oversight. In LEASH: Uncertainty-Aware Weak-to-Strong Escalation and Oversight for Computer-Use Agents (arXiv:2610.22280, 15 Oct 2026), researchers Qingni Wang, Changdae Oh, Yuzhe Yang, Sharon Li, Jiachen Yang, Ang Li, and Xin Eric Wang present a training-free framework that governs escalation and post-handoff oversight using zero-overhead latent uncertainty signals.

LEASH operates across three defensive layers: initial task-level uncertainty estimation for upfront routing, step-level uncertainty monitoring for runtime handoffs, and a post-handoff shadow observer that keeps the black-box strong agent under continuous oversight. Crucially, LEASH estimates uncertainty by comparing the weak agent’s internal hidden states directly against reference representations extracted from its past successful and failed trajectory archives. Because this comparison reuses the latent activations already generated during action sampling, it incurs zero auxiliary forward passes and requires no specialized monitor fine-tuning. Evaluated on the OSWorld desktop benchmark across four lightweight backbones (including EvoCUA-8B, Qwen3.5-35B-A3B, and Nemotron-3-Nano-Omni) paired with Claude Opus 5 as the strong executor, LEASH outperformed random escalation baselines by 6.55 to 11.94 percentage points at matched escalation budgets. When paired with Qwen3.5-35B-A3B, LEASH attained a 73.66% task success rate at $0.4117 per task, matching the performance of always-on Claude Opus 5 (72.90%) while slashing total inference costs by 19.2%.

Why it matters: For autonomous computer-use agents to see practical deployment across operating systems and enterprise tooling, inference costs must drop significantly without compromising safety or success rates. LEASH demonstrates that a model’s internal latent activations provide a sensitive, zero-cost gauge of operational uncertainty. Moreover, by retaining the weak model as an independent runtime monitor after escalation, LEASH establishes a practical mechanism for supervising proprietary black-box agents during critical execution steps.

Caveats: The reliability of latent uncertainty estimation depends heavily on the representativeness and diversity of the historical trajectory bank; anomalous or out-of-distribution system states that do not resemble indexed failure modes can lead to missed escalation triggers. Additionally, because the post-handoff shadow monitor relies on the weak model’s comprehension of the strong model’s actions, the supervisor may fail to detect subtle hallucinations or errors during highly complex, multi-modal desktop sequences.