LEASH: weak-to-strong escalation matches Claude Opus 5 at 19.2% lower cost on OSWorld
xwang_lk · x · 2026-10-17
Simular and UW-Madison propose LEASH, a training-free framework for computer-use agents that lets a weak model act by default and escalates to a strong model only when needed. All three layers — an initial uncertainty observer for early routing, step-level uncertainty for runtime handoff, and a shadow observer for continued oversight of black-box API agents — are fueled by latent uncertainty, reusing activations generated during action creation so no extra forward passes or dedicated monitor training are needed. On OSWorld across four backbones, LEASH beats random escalation by 6.55–11.94 points at matched escalation rates; with Qwen3.5-35B-A3B it hits 73.66% task success at $0.4117/task, matching always-on Claude Opus 5 (72.90%) while cutting inference cost by 19.2%.
More from coding & agent
- Backend migration from Python to Go now takes one engineer just one week — rakyll · 2026-10-17
- Audio founder says Claude one-shot decompiled rival algorithms, moat gone — ChrisGPT · 2026-10-17
- Google's playbook for agent evals: four layers of checks beyond the vibe check — rseroter · 2026-10-17
- AgentForesight: a 7B online auditor that catches multi-agent failures mid-run — jiqizhixin · 2026-10-17
- Dev gives his agent an alarm tool to wake him during overnight /goal runs — zack_overflow · 2026-10-17
- Dev shares multi-agent orchestration recipe: 2-3 lead agents with many focused subagents — MoonL88537 · 2026-10-17