Quick on the Uptake

Research code for the paper "Quick on the Uptake: Eliciting Implicit Intents from Human Demonstrations for Personalized Mobile-Use Agents".

Paper link: https://arxiv.org/abs/2508.08645

🚀 Quick Start

1. Environment Setup

1. Clone the repository:

   git clone https://github.com/MadeAgents/Quick-on-the-Uptake

2. Navigate into the project directory:

   cd Quick-on-the-Uptake

3. MobileIAR dataset link: https://huggingface.co/datasets/wuuuuuz/MobileIAR

   Please download the data and save it as Trajectories.

4. Model link：Warmed-up query rewriter https://huggingface.co/wuuuuuz/IFRAgent.

(Steps 2 to 4 can be skipped as we have already prepared the personalized query and personalized SOP in the json file of the MobileIAR dataset.)

2. Get explicit intent flow and implicit intent flow

python get_intent_flow.py

3. Get explicit intent flow vector database

python get_rag.py

4. Get personalized query and personalized SOP

python rewriting_loop.py

5. Obtain experimental results from the paper

python test_loop_{xxx}.py

Here, xxx refers to different baselines. If you set test_func=test_loop_{xxx}, you will get the experimental results of the baseline. If you set test_func=test_loop_IFRAgent, you will get the experimental results of xxx after being processed by IFRAgent.

📑 Citation

Please cite our paper if you use this toolkit:

@article{wu2025quick,
  title={Quick on the Uptake: Eliciting Implicit Intents from Human Demonstrations for Personalized Mobile-Use Agents},
  author={Wu, Zheng and Huang, Heyuan and Yang, Yanjia and Song, Yuanyi and Lou, Xingyu and Liu, Weiwen and Zhang, Weinan and Wang, Jun and Zhang, Zhuosheng},
  journal={arXiv preprint arXiv:2508.08645},
  year={2025}
}