arxiv:2605.29447

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

Published on May 28

· Submitted by

Tianpeng Bu on Jun 1

Upvote

Authors:

Tianpeng Bu ,

Xin Liu ,

Abstract

GUI agents lack robust error recovery capabilities, which this work addresses through GUI-RobustEval and Robustness-driven Trajectory Synthesis, demonstrating improved performance on real-world benchmarks.

AI-generated summary

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.

View arXiv page View PDF GitHub 4 Add to collection

Community

smallnono

Paper author Paper submitter 1 day ago

We observe that existing GUI agent benchmarks focus on overall success rates or robustness to external perturbations, while real-world failures are predominantly caused by the agent's own policy — and current training data significantly underrepresents planning and progress-perception errors that dominate real-world failures. Motivated by this gap, we propose GUI-RobustEval, a fine-grained benchmark for measuring error awareness and recovery across 11 error types and 4 error depths, and RoTS, a tree-based online synthesis framework that generates 800K error-recovery trajectories by actively exploring failure modes and synthesizing long-horizon recovery data. RoTS-trained models achieve strong performance on OSWorld with substantially smaller degradation under compounding errors compared to existing baselines, demonstrating improved robustness to long-horizon policy-induced failures.