High-priority alert based on source significance and keyword relevance.
Context
What happened: Researchers introduced Vela, a framework that converts anonymized real user behavior records into trainable tasks and trajectories to supervise large language models.
Key numbers: 100,000 users; 100,000 tasks across 91 templates; about 75,000 tasks retained after filtering; fine-tuning Qwen3.6-35B-A3B gained up to 11.5 points on user-related tasks, 6.2 points across
Why it matters: The authors say real-world user behavior is a scalable supervision source for user-centered capabilities and broader LLM intelligence, supporting long-term planning and autonomous agent execution.