Recovering Policy-Induced Errors: Benchmarking and Trajectory Synthesis for Robust GUI Agents

2026-05-28

AI Summary

Existing GUI agents exhibit insufficient robustness in real-world scenarios, particularly in their inability to recover from their own errors. To address this limitation, this work introduces GUI-RobustEval, the first systematic benchmark encompassing diverse error patterns for evaluating GUI agent robustness. Furthermore, we propose RoTS, a robustness-driven tree-based trajectory synthesis framework that actively identifies failure points and generates recovery paths through an extensible tree structure, integrated with large-scale instruction tuning and executable test case generation. Models trained with this framework—RoTS-7B and RoTS-32B—demonstrate substantially enhanced error recovery capabilities, achieving state-of-the-art performance on both GUI-RobustEval and OSWorld. Notably, RoTS-32B attains a success rate of 47.4% and an All-Pass@4 score of 33.8% on OSWorld.

Abstract

While GUI agents have advanced rapidly, they often lack the robustness to recover from their own errors, hindering real-world deployment. To bridge this gap at both the evaluation and data levels, we introduce GUI-RobustEval and propose Robustness-driven Trajectory Synthesis. GUI-RobustEval contains $1,216$ executable test cases that systematically measure error recovery capabilities across a broad and realistic spectrum of error modes. At the data level, RoTS is a scalable synthesis framework that creates $800k$ high-quality data via a tree-based pipeline that proactively discovers diverse error modes and synthesizes corresponding recovery steps. Our two models, RoTS-7B and RoTS-32B, fine-tuned on our dataset, both demonstrate significant gains on GUI-RobustEval and traditional GUI benchmarks. Notably, RoTS-32B achieves state-of-the-art performance on OSWorld, with a $47.4\%$ success rate and a $33.8\%$ All-Pass@4 score, suggesting that improved long-horizon error recovery ability contributes to both robustness and overall performance. Our code is available at https://github.com/AlibabaResearch/RoTS.

Problem

• GUI agents
• error recovery
• robustness
• policy-induced errors
• trajectory synthesis

Innovation

• error recovery
• trajectory synthesis
• GUI agents
• robustness benchmarking
• tree-based data generation