← Back to feed
Papers·1개월 전

OpenMobile: Open-source framework synthesizes mobile agent training data, Qwen3-VL reaches 64.7% on AndroidWorld

OpenMobile: Open-source framework synthesizes mobile agent training data, Qwen3-VL reaches 64.7% on AndroidWorld

OpenMobile, an open-source framework from Nanjing University, synthesizes high-quality task instructions and agent trajectories for mobile agents. It uses a scalable task synthesis pipeline with global environment memory and a policy-switching strategy (learner/expert) to capture error-recovery data. Fine-tuned Qwen2.5-VL and Qwen3-VL achieve 51.7% and 64.7% on AndroidWorld, surpassing existing open-data approaches, with analysis confirming gains come from functionality coverage rather than overfitting.

  • #mobile-agent
  • #vision-language-model
  • #open-source
  • #data-synthesis
  • #androidworld
Kanzhi Cheng
원문 보기 →

Comments

— 첫 댓글을 남겨보세요 —