LEASH: training-free uncertainty framework cuts computer-use agent cost via weak-to-strong escalation
Qingni Wang · hf · 2026-10-17
LEASH, a training-free framework from Similar and UW–Madison, tackles the cost problem of computer-use agents where running a frontier model at every step is expensive but a small open-weight model alone often fails.
It uses weak-to-strong escalation: a weak agent acts by default and hands control to a strong one only when needed, with continued oversight afterward. Since the strong agent is a black-box API, LEASH uses latent uncertainty as fuel for three layers: (1) task-level uncertainty for early routing, (2) step-level uncertainty for runtime escalation, and (3) a shadow observer monitoring the strong black-box agent after handoff. Uncertainty is estimated by comparing the weak model's hidden states against those from its past successful and failed trajectories.
Result: reduced cost while maintaining high success rates.
Related event: LEASH: Training-Free Framework Cuts Computer-Use Agent Costs by 19.2%(3 posts)→
More from coding & agent
- Solo dev builds Waifu Weather iPhone app with Claude Opus and Sonnet, opens TestFlight beta — Atomosic · 2026-10-17
- Agent ports a Rails app to Jai with one prompt: 'things move fast' — lucasmeijer · 2026-10-17
- From 'git push heroku master' to telling Codex: keep my app deployed and healthy — alex_frantic · 2026-10-17
- AI finally configures a Home Assistant setup the author dodged for 3 months — haydendevs · 2026-10-17
- Delegating to dumber models to save tokens is a bad trade, argues heavy AI user — brandon_galang · 2026-10-17
- Dev builds Subway Surfers-style flying Tesla Roadster game with Gemini 4 Argon — ammaar · 2026-10-17