VISTA Improves GUI Grounding with View-Consistent Self-Verified Training

A new research paper presents VISTA, which stands for View-Consistent Self-Verified Training, a novel framework designed to improve GUI grounding. This method addresses limitations in existing Group Relative Policy Optimization (GRPO) by generating comparison groups from multiple geometrically varied but semantically equivalent views of a single GUI instance. Each view ensures the target element remains visible and its bounding box is precisely remapped. VISTA also integrates a self-verified cross-view anchor, an optimized oracle answer, to stabilize the generation of short coordinates. This anchor is activated only when the model achieves a maximum-reward rollout, preventing the reinforcement learning process from devolving into simple imitation. The framework has demonstrated consistent improvements in grounding accuracy across five GUI-grounding benchmarks and various Qwen backbones, showing enhanced robustness and reduced prediction instability.