search

mnotgod96 / AppAgent

AppAgent: Multimodal Agents as Smartphone Users, an LLM-based multimodal agent framework designed to operate smartphone apps.
https://appagent-official.github.io/
MIT License
4.84k stars 511 forks source link

On reward model #36

Open mazzzystar opened 7 months ago

mazzzystar commented 7 months ago

Hi, thank you for making AppAgent open source. It's a fascinating concept.

I've read the paper and came across this paragraph:

Reward: To provide a more fine-grained measurement, we developed a reward model to assess performance. For each task within an app, we scored different UI pages. The closer the UI page was to the objective, the higher the score it received. This means that even if the agent failed to complete the task, it would still receive credit based on its final state.

However, I couldn't find the code for the reward model in this repository. I'm interested in understanding how to train and deploy the reward model. Many thanks!