Reproduce “Vela: User Behavior as Supervision Sparks LLM Intelligence” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yang Xiao, Yusong Sun, Haoyi Wu, Wenyang Hui, Weiqiang Hang, Luqing Wang, Zejian Wang, Zirui Tu and 12 more
Vela turns long behavior histories into predictive and rubric-scored tasks, using later actions as targets and visible behavior as evidence. Fine-tuning Qwen3.6-35B-A3B as Vela-35B-A3B raises VelaBench from 51.5 to 63.0, a gain of 11.5 points. Across four user-related benchmarks, the mean rises from 41.3 to 47.5; results use anonymized histories from one social-media platform.
Key insights
- Subsequent user actions can supervise both forecasts of what a person may do and evidence-grounded explanations or summaries of that person.
- Training on behavior histories improved selected general benchmarks as well as personalization measures in these experiments, suggesting the benefits were not limited to user-specific tasks.
- A Vela-trained checkpoint also served as a stronger starting point for the tested GSPO reinforcement-learning runs than the base model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.