$ cat wiki/papers/2026/2610.22280-leash.md
LEASH
TL;DR
LEASH uses a weak model's latent uncertainty to decide when to hand a computer-use task to a stronger model, then continues monitoring after the handoff. It targets Model Routing within a task. (source)
Authors & Org
Qingni Wang, Changdae Oh, Yuzhe Yang, Sharon Li, Jiachen Yang, Ang Li, Xin Eric Wang. Affiliations are not established by the abstract record read here. First submitted 2026-10-15. (source)
Method
The training-free framework combines initial task routing, step-level escalation, and a shadow observer that oversees the strong black-box agent. It compares the weak model's hidden states with states from earlier successful and failed trajectories. During weak-agent execution, uncertainty estimation reuses action-generation activations, requiring no extra forward pass or dedicated monitor training; that saving is not stated for every stage of the system. (source)
Results
On OSWorld across four model backbones, the authors report gains of 6.55–11.94 points over random escalation at matched escalation rates. With Qwen3.5-35B-A3B, LEASH reports 73.66% task success at $0.4117 per task, compared with always-on Claude Opus 5 at 72.90%, while lowering inference cost by 19.2%. These are author-reported results, not an independent replication. (source)
Significance
For Agents (LLM Agents), the useful distinction is between choosing a stronger model and supervising it after selection. The random-escalation comparison tests whether the uncertainty signal helps beyond spending the same handoff budget. (source)
Open Questions
The abstract does not state confidence intervals, the size and acquisition cost of the trajectory reference set, or the shadow observer's separate overhead. Full paper and code were not reviewed, so robustness under new applications and the complete cost accounting remain open here. (source)
Cite
“LEASH: Uncertainty-Aware Weak-to-Strong Escalation and Oversight for Computer-Use Agents.” arXiv:2610.22280. This note uses the abstract and metadata in the snapshot.