HalluClear Framework Introduced to Diagnose and Mitigate Hallucinations in GUI Agents

A new research framework called HalluClear addresses the persistent problem of ungrounded hallucinations in GUI agents, which frequently cause cascading failures during real-world deployment. Unlike broader visual language model domains, the GUI agent field has lacked specialized tools for fine-grained diagnosis, reliable evaluation, and targeted mitigation of these errors. HalluClear fills this gap as a comprehensive suite designed to complement computation-intensive scaling approaches. The framework includes a GUI-specific hallucination taxonomy developed through empirical failure analysis. It also features a calibrated three-stage evaluation workflow that improves the reliability of VLM-as-a-judge assessments via expert-annotated benchmarking and ensemble credibility estimation. Additionally, HalluClear incorporates a mitigation scheme based on closed-loop structured reasoning. This scheme enables lightweight continual post-training with cold-start initialization, applicable to both generalist and GUI-specialist agents. The research was announced on arXiv under the identifier arXiv:2604.17284v1.