AllSpark-Research/Vela

你做 personal agent，最难的一类请求大概是这样：用户只说了半句话，答案藏在几个月前的搜索和最近的反馈里。 Vela 把真实行为记录变成监督信号。约 10 万个匿名行为主体，六类行为，91 种场景模板，拼出 10 万条候选任务，留下 7.5 万条训练样本。论文把它叫「LLM 智能的用户维度」。…

A research spotlight from Zhihu’s 「AI Researcher Club」 Vela: User Behavior Can Teach LLMs More Than Personalization Training on user behavior improved Vela across all 10 benchmarks tested, including…

9/10 A stronger starting point for RL. With the same RL setup, Vela-initialized models score 7.7 points higher on both VelaBench and ProfileSummary than models initialized from the base. The Vela…

10/11 A stronger starting point for RL. With the same RL setup, Vela-initialized models score 7.7 points higher on both VelaBench and ProfileSummary than models initialized from the base. The Vela…

Vela: learning LLMs from real user behavior From the anonymized behavior of 100K users, Vela generates 75K tasks to fine-tune a 35B model, improving both user-related and general benchmarks.