Large language models are increasingly deployed in settings that depend on users' long-term behavioral histories, from personal assistants to recommendation and advertising systems. These applications require what the authors call the user dimension of LLM intelligence: inferring a person's preferences, intent, and likely future actions from weeks to years of behavior data.
Current models, however, are typically trained on static text or synthetic data, neither of which reflects how user behavior evolves over time. The paper argues that real-world behavior records, already available at scale on content platforms and in personal-assistant agents, offer a natural supervisory signal.
The proposed framework, Vela, converts real user behavior data into trainable tasks and trajectories. From anonymized behavior records of 100,000 users, it generates 100,000 tasks across 91 real-world task templates. Difficulty-based filtering and trajectory acceptance criteria retain approximately 75,000 tasks with selected training trajectories.
Fine-tuning Qwen3.6-35B-A3B on these trajectories improved performance over the base model by up to 11.5 points on user-related tasks and by an average of 4 points on general benchmarks. The authors also report that training on more user behavior data further improves user-related performance, and that reinforcement learning after fine-tuning outperforms reinforcement learning from the base model by up to 7.7 points.
According to the first author, Yang Xiao, the fine-tuned model improved on all ten evaluated benchmarks, with gains of 11.5 points on VelaBench, an average of 6.2 points across four independent user-related benchmarks, and 4.0 points across five general benchmarks. Behavior records were anonymized before task construction, and their use was platform-authorized and underwent privacy and ethics review.
The authors conclude that real-world user behavior can serve as a scalable source of supervision for both user-centered capabilities and the broader intelligence of LLMs.