LEASH: uncertainty-aware weak-to-strong escalation matches Claude Opus 5 at 19% lower cost
SharonYixuanLi · x · 2026-10-24
Simular and UW-Madison propose LEASH, a training-free weak-to-strong escalation framework for computer-use agents: a lightweight model acts by default and hands off to a frontier model only when its latent uncertainty signals the need, with a shadow observer continuing to oversee the black-box strong agent.
Three layers, all fueled by latent uncertainty compared against past successful/failed trajectories:
- task-level uncertainty for initial routing;
- step-level uncertainty for runtime escalation — the signal reuses activations already produced, so no extra forward pass or monitor training;
- a shadow observer for post-handoff oversight without access to the strong agent's internals.
Results on OSWorld: LEASH beats random escalation by 6.55–11.94 points at matched rates. With Qwen3.5-35B-A3B as the weak agent it reaches 73.66% task success at $0.4117/task, matching always-on Claude Opus 5 (72.90%, $0.51) while cutting inference cost by 19.2%, and gains 9.1 points over random routing at equal cost.
More from coding & agent
- 20-year AI veteran curates 19 free open-source repos for building AI agents — MaryamMiradi · 2026-10-24
- Microsoft retires Fabric data agent custom tool in Foundry, pivots to Fabric IQ — adnan_hashmi · 2026-10-24
- A 2-person AI team built its own video tool and dropped Zoom entirely — danshipper · 2026-10-24
- Agent posting safety: publish, fetch as a stranger, then assert on what's actually there — petrusenko_max · 2026-10-24
- 100% test coverage, zero bugs found: why this dev added live-post verification — petrusenko_max · 2026-10-24
- Claude approved its own broken 3D screenshots — how a 480Hz verification gate fixed agent self-grading — youthanasia669 · 2026-10-24