Step-level Optimization for Efficient Computer-use Agents
Proposed event-driven step-level cascade framework improves efficiency, achieving 58.2% success rate.
Key Findings
Methodology
The study introduces an event-driven step-level cascade framework to optimize computer-use agents' efficiency. The framework uses lightweight monitors to detect risks and selectively invoke stronger models. It includes a Stuck Monitor for progress degradation detection and a Milestone Monitor for semantic checkpoint identification.
Key Results
- On OSWorld, EvoCUA-8B + Kimi K2.5 achieved a 58.2% success rate, reducing costs by 61.4%.
- On WebArena, gpt-oss-20b + GPT-5.2 achieved a 57.8% success rate, significantly reducing costs and latency.
- Enabling both Stuck and Milestone detectors resulted in significant performance improvements, covering different failure modes.
Significance
This study significantly enhances the efficiency of computer-use agents, addressing the computational resource waste in long-horizon GUI tasks. Through intelligent compute allocation, it reduces costs and latency, advancing software automation.
Technical Contribution
Introduces a modular, deployment-oriented framework that can be layered on existing agents without changing architecture or retraining models. Provides a new compute allocation strategy, reducing large model usage frequency.
Novelty
First to implement step-level cascade optimization in computer-use agents, providing finer state-dependent control compared to traditional query-level cascades.
Limitations
- Framework relies on monitor accuracy, which may fail in complex tasks.
- Stuck detection may incorrectly trigger strong model invocation in some cases.
- Milestone detection needs further optimization to reduce false positives.
Future Work
Future directions include optimizing monitor accuracy and stability, and extending the framework to support more task types and environments.
AI Executive Summary
Computer-use agents hold great potential in software automation, but existing solutions are inefficient for long-horizon tasks. To address this, the study proposes an event-driven step-level cascade framework that uses lightweight monitors to detect risks and selectively invoke stronger models. The framework includes a Stuck Monitor for progress degradation detection and a Milestone Monitor for semantic checkpoint identification. Experimental results show that the framework significantly improves success rates on OSWorld and WebArena while reducing costs and latency. This study provides new insights for the software automation field, advancing efficient computer-use agents.
Deep Analysis
Background
Computer-use agents can interact directly with graphical user interfaces, offering a path to general software automation. However, existing agents are inefficient for long-horizon tasks, mainly due to invoking large multimodal models at every step, leading to computational resource waste.
Core Problem
Uneven compute resource allocation in long-horizon GUI tasks leads to inefficiency. Many steps are routine and can be reliably handled by smaller policies, while errors tend to concentrate at a few high-risk moments.
Innovation
Proposes an event-driven step-level cascade framework that uses lightweight monitors to detect risks and selectively invoke stronger models. Includes a Stuck Monitor for progress degradation detection and a Milestone Monitor for semantic checkpoint identification.
Methodology
- �� Use lightweight monitors to detect risks
- �� Stuck Monitor detects progress degradation
- �� Milestone Monitor identifies semantic checkpoints
- �� Selectively invoke stronger models for recovery
Experiments
Experiments conducted on OSWorld and WebArena to evaluate the framework's efficiency and success rates. Various combinations of small and large models were used to compare performance across different configurations.
Results
On OSWorld, EvoCUA-8B + Kimi K2.5 achieved a 58.2% success rate, reducing costs by 61.4%. On WebArena, gpt-oss-20b + GPT-5.2 achieved a 57.8% success rate, significantly reducing costs and latency.
Applications
The framework can be used to improve efficiency in various software automation tasks, especially in scenarios requiring long-term interaction, such as desktop applications and web navigation.
Limitations & Outlook
Framework relies on monitor accuracy, which may fail in complex tasks. Stuck detection may incorrectly trigger strong model invocation, and milestone detection needs further optimization to reduce false positives.
Plain Language Accessible to non-experts
Imagine you're cooking in a kitchen, sometimes you need a small knife for chopping vegetables, and sometimes a big knife for cutting meat. Our framework is like a smart chef who knows when to use the small knife and when to use the big one, saving time and reducing kitchen chaos.
ELI14 Explained like you're 14
Hey there! Imagine you're playing a game, sometimes you need to use small skills, and sometimes you need to use big moves. Our framework is like a smart game character who knows when to use small skills and when to use big moves, helping you win the game and save energy!
Glossary
Event-driven
A mechanism that triggers actions based on events.
Used to detect risks and trigger model invocation.
Step-level cascade
An optimization strategy that selects models at each step.
Used to improve compute efficiency.
Stuck Monitor
A component that detects progress degradation.
Used to trigger strong model invocation.
Milestone Monitor
A component that identifies semantic checkpoints.
Used to verify task progress.
Semantic drift
The phenomenon where an agent deviates from the user's goal.
Needs detection by the Milestone Monitor.
Open Questions Unanswered questions from this research
- 1 How to improve monitor accuracy to reduce false positives?
- 2 How to ensure framework stability in complex tasks?
Applications
Immediate Applications
Desktop Application Optimization
Improve task efficiency in desktop applications, reducing compute costs.
Long-term Vision
Fully Automated Software Systems
Achieve more efficient automated software systems, reducing human intervention.
Abstract
Computer-use agents provide a promising path toward general software automation because they can interact directly with arbitrary graphical user interfaces instead of relying on brittle, application-specific integrations. Despite recent advances in benchmark performance, strong computer-use agents remain expensive and slow in practice, since most systems invoke large multimodal models at nearly every interaction step. We argue that this uniform allocation of compute is fundamentally inefficient for long-horizon GUI tasks. Such trajectories are highly heterogeneous: many steps are routine and can be handled reliably by smaller, cheaper policies, while errors tend to concentrate at a relatively small number of high-risk moments. Across computer-use benchmarks, these failures repeatedly take two forms: progress stalls, where the agent loops, repeats ineffective actions, or fails to make meaningful progress, and silent semantic drift, where the agent continues taking locally plausible actions after already deviating from the user's true goal. To address this inefficiency, we propose an event-driven, step-level cascade for computer-use agents that runs a small policy by default and escalates to a stronger model only when lightweight learned monitors detect elevated risk. Our framework combines two complementary signals: a Stuck Monitor that detects degraded progress from recent reasoning-action history and triggers recovery, and a Milestone Monitor that identifies semantically meaningful checkpoints where sparse verification is most informative for catching drift. This design turns always-on frontier-model inference into adaptive, on-demand compute allocation over the course of an evolving interaction. The framework is modular and deployment-oriented: it can be layered on top of existing computer-use agents without changing the underlying agent architecture or retraining the large model.