AI Research

Vela: User Behavior as Supervision Sparks LLM Intelligence | AI Learning Platform

BenchmarkAgentFine-tuningRLHF

Vela: User Behavior as Supervision Sparks LLM Intelligence

Yang Xiao, Yusong Sun, Haoyi Wu et al.1 min readHugging FaceOct 10, 2026

Key Takeaway for Students

Large language models are increasingly deployed in applications that act on users' long-term behavior histories, from personal assistants to recommendation and advertising systems. These applications demand the user dimension of LLM intelligence: inferring a person's preferences, intent, and likely future actions from weeks to years of behavior data. Yet current models are trai

outperforms reinforcement learning from the base model by up to 7

Reading in Plain English mode — toggle below to switch

This work introduces a novel approach to training LLMs by leveraging real user behavior as a supervisory signal, moving beyond static or synthetic data. For ML engineers, this opens up new avenues for developing more personalized and context-aware AI applications, particularly in areas like personal assistants and recommendation engines.

Read original on Hugging Face

Career Insight

high demand

Why it matters

Skills to learn

Designing data pipelines for processing long-term user behavior dataImplementing custom training loops for LLMs with dynamic supervisory signalsDeveloping strategies for anonymizing and handling sensitive user data at scaleEvaluating LLM performance on tasks requiring user preference inference and predictionApplying RLHF principles to user interaction trajectories

Interview questions

Describe the challenges of using real-world, long-term user behavior data for LLM supervision compared to static text datasets. How would you address data sparsity and noise?
Vela proposes converting user behavior into 'trainable tasks and trajectories'. Can you elaborate on the types of tasks and trajectory representations you might design for a personal assistant LLM, and how you would formulate the loss function?
Design a system architecture for a large-scale LLM training pipeline that continuously incorporates new user behavior data to update the model. Consider data ingestion, anonymization, task generation, and model retraining/deployment strategies.

Relevant for ~20 months · generated by the enrichment pipeline