Vela: User Behavior as Supervision Sparks LLM Intelligence

Yang Xiao, Yusong Sun, Haoyi Wu, Wenyang Hui, Weiqiang Hang, Luqing Wang et al.

arXiv 2610.14709 22 pages Read the paper

The short version

The problem: Language models often do not know what one particular person likes or is likely to do, especially when that must be inferred from months or years of actions. This makes personalization and user-facing recommendations unreliable.
The idea: Vela teaches models from what people actually did, rather than relying only on written examples or invented scenarios. It turns a person’s earlier activity into questions—such as what they might buy next—and uses later activity as the answer, while also asking for evidence-based summaries and explanations.
What it got: The trained model improved by 11.5 points on the paper’s main test of user-related tasks, and by an average of 4 points on general tests. The authors also found that training from this user-behavior foundation led to a 7.7-point advantage after further reinforcement learning on two user-related tests.

Modules from this paper

User Behavior as a Foundation for Reinforcement Learning

A base model can struggle to learn user-facing tasks from reinforcement learning alone; Vela tests whether supervised training on real user behavior gives that same RL setup a stronger starting point.

reinforcement-learningsupervised-fine-tuninguser-modelingpost-training

Vela: Selecting Tasks and Trajectories by Difficulty

Behavior-derived training data can include tasks that are too easy and generated answers that are weak; Vela filters for intermediate-difficulty tasks and keeps only trajectories that beat a task-specific score threshold.

data-curationinstruction-tuningsynthetic-datatraining

Vela: User Behavior as Supervision

Training models from user histories is hard to score consistently; Vela turns earlier behavior into evidence and later behavior into prediction targets, alongside rubric-judged summaries and explanations.

supervisionlanguage-modelsuser-behaviortask-design