Failure-Aware Dual-Flow Control for Computer-Use Agents

ACL ARR 2026 January Submission10679 Authors

Keywords: Computer-use agents, agent, LLM, failure recovery

Abstract: Computer-use agents (CUAs) operate by directly interacting with graphical user interfaces in real operating system environments, where execution failures frequently prevent progress. When actions fail to induce observable state changes, existing agents often continue conditioning on unchanged observations and repeatedly generate similar decisions, exhausting execution budgets without effective recovery. We frame this behavior as a control problem: current agentic frameworks lack explicit mechanisms to regulate how decision-making should adapt under execution failure. To address this gap, we propose Failure-aware Dual-flow Control (FDC), a control framework that explicitly separates standard execution from adaptive recovery and regulates transitions between them using action-grounded failure signals. FDC integrates action-grounded dual failure evaluation with a hierarchical adaptive recovery mechanism, enabling targeted decision revision only upon detected failure. Evaluated on OSWorld, a large-scale benchmark for real OS interaction, FDC consistently improves success rates across both open-weight and closed-source backbones while incurring substantially lower inference-time cost than scaling-based alternatives. These results demonstrate that explicit failure-aware control is critical for robust and efficient computer-use agents.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: AI / LLM Agents

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 10679