LEASH helps AI agents know when to ask for help

Asking for help at the right time

LEASH is a new method that makes AI agents smarter about when to call in a more powerful system. Computer-use agents are programs that can click buttons, type text, and navigate software on their own. They are useful for long tasks, but running a top-tier model for every single step is expensive, and a small open-source model often fails on its own. Weak-to-strong escalation is the idea of using a weaker agent by default and handing control to a stronger one only when needed, but the hard part has always been knowing exactly when to escalate.

The researchers show that the weak agent's own internal signals can answer that question. They compare the agent's hidden states, or the internal representations produced during each decision, with memories of past successful and failed runs. If a current state looks more like the failures, the system treats it as high uncertainty and calls the stronger model. Because the estimator reuses activations already produced during action generation, no extra forward pass or separate monitor training is required.

Three layers of oversight

LEASH works in three stages. The first stage estimates initial task uncertainty to route hard problems straight to the strong agent. The second checks uncertainty at every step while the weak agent is working, escalating dynamically if things go off track. After the strong agent takes over, the third layer, a shadow observer, continues to watch it even though most commercial APIs hide their internal details from outside monitoring.

All three layers rely on latent uncertainty drawn from the weak model's hidden states. This keeps the system training-free and cheap to run.

Real results on real benchmarks

Tests on OSWorld, a benchmark for AI agents that complete computer tasks, used four contemporary model backbones. Across all of them, LEASH improved over standalone weak agents and beat random escalation by 6.55 to 11.94 points at the same escalation rates ¹.

With Qwen3.5-35B-A3B, LEASH reached 73.66 percent task success at about $0.41 per task, while always-running Claude Opus 5 scored 72.90 percent at higher cost ¹. The weak-to-strong approach matched the expensive model and cut inference cost by roughly 19 percent.

Why this matters

As AI agents take on longer, more complex workflows, paying for a frontier model at every click becomes a real bottleneck. LEASH shows that uncertainty can be estimated from existing model signals without adding heavy infrastructure. That makes escalation practical for systems that already depend on closed APIs, where internal steps are otherwise invisible.

The work does not claim to solve every oversight problem. Black-box APIs still hide their reasoning, and the shadow observer described here works with latent comparisons rather than direct access to the strong model's internals. Still, the results suggest that better routing, not just bigger models, may be the more scalable path forward.

¹

Key facts

• LEASH is a training-free framework that escalates from weak to strong AI agents based on latent uncertainty signals.
• It uses three layers: task-level routing, step-level escalation, and a shadow observer after handoff.
• Tests on OSWorld show LEASH outperforms random escalation by 6.55 to 11.94 points at matched escalation rates.
• With Qwen3.5-35B-A3B, LEASH reaches 73.66 percent success at $0.41 per task, matching Claude Opus 5 at 19 percent lower cost.

Sources

1. Paper page - LEASH: Uncertainty-Aware Weak-to-Strong Escalation and Oversight for Computer-Use Agents