SimuLang을 소개합니다: 데스크톱 전체를 위한 Playwright
캘리포니아주 팔로알토 • 2026년 8월 28일 • 작성: Simular 팀
Computer-use agents increasingly tackle long-horizon tasks, where running a frontier model at every step is expensive and a small open-weight model alone often fails.
Weak-to-strong escalation offers a middle ground: a weak agent acts by default and hands control to a strong one only when needed. However, practical escalation faces a dilemma:
- Deciding when to escalate: This requires an uncertainty signal cheap enough to check at every step.
- Post-handoff risk: After handoff, the strong agent can still go astray, and a black-box API exposes no internals to monitor.
To solve this, Simular’s research team proposes LEASH, a training-free framework that unleashes the strong agent only when the weak agent's latent uncertainty signals it is needed, and keeps it on a leash afterward.
결과
순수 모델의 경우 성공률이 높을수록 비용도 증가합니다. Sai는 이러한 곡선을 깨뜨렸습니다. 차트에서 가장 높은 점수를 기록하면서도 경쟁 모델 대비 3분의 2 미만의 비용으로 운영됩니다.
- 효율적인 캐싱 및 메모리: Sai는 적응형 요약을 통해 입력 크기를 제한하고, 요약 전까지는 최대한 일정한 프롬프트 접두사를 유지합니다. 코드를 통한 계획 수립으로 Sai는 작업 전체 실행 과정에서 중요한 정보를 런타임 메모리에 유지하고 액세스할 수 있습니다.
- 모델 오케스트레이션: Sai는 UI 요소 위치 파악, 순수 추론, 검증을 위해 전문 모델과 인터페이스를 호출합니다. 이를 통해 전체 작업 맥락이 필요하지 않은 단계에서는 비용이 많이 드는 모델 사용을 피합니다.
- Zero extra forward pass: During weak-agent execution, runtime uncertainty reuses activations already produced during action generation, requiring no extra forward pass and no dedicated monitor training.
Because it reuses existing activations, latent uncertainty provides a signal cheap enough to check at every step.
The Three Layers of LEASH
LEASH coordinates execution across three layers, all powered by latent uncertainty:
- Initial Task Uncertainty (Early Routing): Evaluates latent uncertainty at the beginning of the task to enable early routing directly to the strong agent when needed.
- Step-Level Uncertainty (Runtime Escalation): Monitors latent uncertainty at every step during weak-agent execution, escalating control to the strong agent when high uncertainty is detected.
- Shadow Observer (Continued Black-Box Oversight): After escalation, the weak agent acts as a shadow observer to provide continued oversight of the strong black-box agent, overcoming the limitation that black-box APIs expose no internals to monitor.
Evaluation on OSWorld
Experiments on the OSWorld benchmark across four contemporary weak-agent backbones demonstrate:
- Consistent improvements: LEASH consistently improves over standalone weak agents across all four evaluated backbones.
- Outperforming random escalation: At matched escalation rates, LEASH outperforms random escalation by 6.55 to 11.94 points, validating the effectiveness of the latent uncertainty signal.
- Matching frontier performance at lower cost: With Qwen3.5-35B-A3B as the weak backbone, LEASH achieves 73.66% task success at $0.4117 per task, matching always-on Claude Opus 5 (72.90%) while reducing inference cost by 19.2%.
"Each model provider wants you to use only their model. Sai frees you from such lock-in: it calibrates the different strengths and weaknesses of various models, to deliver greatest efficiency," said Eric Wang, head of research at Simular.
LEASH demonstrates this principle by combining the efficiency of weak open-weight agents with the capability of strong frontier models, using cheap latent uncertainty to route and supervise execution without lock-in.
For more details, read the paper here: https://arxiv.org/abs/2610.22280
