The arXiv abstract for "Vela: User Behavior as Supervision Sparks LLM Intelligence" opens by noting that large language models are being deployed in applications that act on users' long-term behavior histories, such as personal assistants, recommendation systems, and advertising. These applications, the abstract argues, demand a "user dimension" — but the sentence cuts off there.

Because the abstract is truncated, the full proposal is not available in the source. The title indicates that the authors treat user behavior as a supervision signal to spark LLM intelligence, but the specific mechanism, training setup, and any experimental evidence are not described in the text provided.

This digest can therefore only report the motivation: the paper identifies a gap in how LLMs handle long-term user context and proposes behavior-based supervision as a direction. Readers interested in the actual method will need to consult the full paper.