Vela: User Behavior as Supervision Sparks LLM Intelligence
Large language models are increasingly deployed in applications that act on users' long-term behavior histories, from personal assistants to recommendation and advertising systems.
This article introduces Vela, a framework that leverages real-world user behavior data as a supervisory signal to enhance large language models (LLMs). The approach aims to improve user-centric capabilities and broader LLM intelligence by analyzing long-term behavior histories.
What happened
The article presents Vela, a framework that processes 100,000 users' anonymized behavior records to generate 100,000 tasks. These tasks are used to fine-tune LLMs, demonstrating measurable performance improvements on user-related and general benchmarks.
Why it matters
User behavior data offers a practical, scalable alternative to static or synthetic training data, enabling LLMs to better adapt to real-world user preferences and interactions, which is critical for applications like personal assistants and recommendation systems.
- User behavior data provides a valuable, scalable supervision source for LLMs.
- Fine-tuning with Vela's task generation improves both user-specific and general model performance.
- Reinforcement learning benefits from user behavior-based pre-training.
- Long-term behavior analysis enhances LLMs' ability to infer user intent and preferences.
The bigger picture
The approach demonstrates that real-world user data can significantly improve LLM performance for user-centric applications, potentially influencing future model training strategies and data utilization in AI systems.
Vela establishes a novel method for leveraging real-world user behavior data to train LLMs, showing measurable performance gains and highlighting the potential of behavior-based supervision in advancing AI systems that interact with humans over time.
Technical details
- Vela's task generation process uses difficulty-based filtering and trajectory acceptance criteria.
- The framework employs Qwen3.6-35B-A3B as the base model for fine-tuning.
- Reinforcement learning is applied post-fine-tuning to further enhance performance.
- 91 real-world task templates are used to structure user behavior data.
FAQ
What is Vela's primary purpose?
Vela is a framework that converts real user behavior data into trainable tasks to improve LLMs' ability to understand and predict user preferences and actions.
How does Vela differ from traditional LLM training methods?
Unlike static text or synthetic data, Vela uses longitudinal user behavior records to create dynamic, real-world training trajectories for models.
What performance improvements did Vela achieve?
Fine-tuning with Vela's data improved user-related task performance by up to 11.5 points and general benchmarks by 4 points, with further gains from reinforcement learning.
What types of applications could benefit from this research?
Personal assistants, recommendation systems, and advertising platforms that rely on understanding user behavior over time.
Is user data anonymized in Vela's approach?
Yes, the behavior records used in the study are anonymized to protect user privacy.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by arXiv - cs.CL (NLP) on Oct 14, 2026.
Comments
Sign in to join the discussion.