Vela：以用户行为作为监督激发大模型智能
原标题：Paper page - Vela: User Behavior as Supervision Sparks LLM Intelligence
AI 摘要
该论文提出 Vela 框架，将真实用户行为数据转化为可训练任务与轨迹，用于监督大模型训练。作者从 10 万名用户的匿名行为记录中生成覆盖 91 个真实任务模板的 10 万个任务，经难度过滤与轨迹筛选后保留约 7.5 万个训练样本。在 Qwen3.6-35B-A3B 上微调后，用户相关任务最高提升 11.5 分，通用基准平均提升 4 分，且微调后强化学习比从基座模型直接强化学习最高提升 7.7 分。
正文节选
Vela: User Behavior as Supervision Sparks LLM Intelligence Abstract Large language models are increasingly deployed in applications that act on users' long-term behavior histories, from personal assistants to recommendation and advertising systems. These applications demand the user dimension of LLM intelligence: inferring a person's preferences, intent, and likely future actions from weeks to years of behavior data. Yet current models are trained on static text or synthetic data, neither of whi