PPP-Agent is an open-source framework for training LLM agents that are not only productive (task success) but also proactive (ask essential clarifying questions) and personalized (adapt to diverse user preferences). It includes UserVille, an interactive environment that turns existing agent benchmarks into multi-turn, preference-aware simulations.
Training Proactive and Personalized LLM Agents
Author: Weiwei Sun, Xuhui Zhou, Weihua Du, Xingyao Wang, Sean Welleck, Graham Neubig, Maarten Sap, Yiming Yang
https://arxiv.org/pdf/2511.02208
- UserVille: converts precise prompts into vague ones and simulates users with 20 configurable preferences.
- PPP RL: multi-objective RL optimizing Productivity, Proactivity, Personalization jointly.
- Plug-and-Play Tools: SWE (SWE-Bench & SWE-Gym) and Deep-Research (BrowseComp+) scaffolds.
- User-Centric Metrics: effort-based proactivity + preference-following personalization.
- Generalization: transfers to unseen preferences, simulators, and downstream tasks.
If you find this work useful, please consider citing our paper:
@article{sun2025pppagent,
title={Training Proactive and Personalized LLM Agents},
author={Sun, Weiwei and Zhou, Xuhui and Du, Weihua and Wang, Xingyao and Welleck, Sean and Neubig, Graham and Sap, Maarten and Yang, Yiming},
journal={arXiv preprint arXiv:2511.02208},
year={2025},
url={https://arxiv.org/abs/2511.02208}
}This project is licensed under the Apache License 2.0 - see the LICENSE file for details.