Researchers propose a new paradigm for training language models (LLMs) to be proactive and personalized collaborators. They introduce UserVille, an interactive environment with simulated users that provides feedback on LLM performance. A multi-objective reinforcement learning framework is used to optimize three dimensions of collaborative AI: productivity, proactivity, and personalization. The approach outperforms strong LLM baselines by 16.7 points on two real-world tasks, and a user study highlights the importance of user-centric feedback for training effective collaborators. --- Why it matters: This work matters because it addresses the limitations of current AI agents, which are primarily optimized for isolated task completion. By developing proactive and personalized LLMs, researchers can create more effective and efficient collaboration tools that adapt to human needs and preferences. Source: https://arxiv.org/abs/2511.02208