Small GUI agents learn to plan via autonomous exploration + hindsight: 7B model beats Qwen2.5-VL-32B

arXiv·medium signalXBlueskyLinkedInCopy link

A new paper (arXiv:2606.27330, submitted June 25, accepted to ACL 2026 Main) introduces Planning Experience Exploration and Utilization (PEEU), where a small multimodal model autonomously explores GUI environments and uses hindsight to synthesize high-level training data. A 7B model reaches 30.6% accuracy, surpassing the much larger Qwen2.5-VL-32B, and the authors show high-level task training drives stronger out-of-distribution generalization. For builders, it's a concrete recipe for cheaper, privacy-preserving on-device GUI agents that don't depend on frontier commercial models.