Some questions about agent

[bianliuyang]

Hey guys!
Thank you for your excellent work. I have some questions. In the mobile-agent-v3 framework mentioned in the paper, there are multiple agents. In practice, do these agents use a unified model scheduling or do they fine-tune multiple different models based on different data?

[junyangwang0410]

Hello. Mobile-Agent-v3 is only powered by GUI-Owl. All agents are based on the unified GUI-Owl.

[bianliuyang]

Hello. Mobile-Agent-v3 is only powered by GUI-Owl. All agents are based on the unified GUI-Owl.

Thanks for your reply. I have another question. In the SFT experiments you conducted before the reinforcement learning experiments, did the GUI-grounding data contain thought chains? I have always been wondering whether the GUI-grounding data in the SFT stage (for example, question: what is the position of the element corresponding to the command "network and internet settings", answer: [x,y]) requires thought chains.

[zhuzq-iziy]

Hello. Mobile-Agent-v3 is only powered by GUI-Owl. All agents are based on the unified GUI-Owl.

Thanks for your reply. I have another question. In the SFT experiments you conducted before the reinforcement learning experiments, did the GUI-grounding data contain thought chains? I have always been wondering whether the GUI-grounding data in the SFT stage (for example, question: what is the position of the element corresponding to the command "network and internet settings", answer: [x,y]) requires thought chains.

Hello, our GUI grounding does not incorporate chain-of-thought reasoning. Training solely on click coordinates alone is sufficient to achieve promising results.