Query as Anchor: Scenario-Adaptive User Representation via Large Language Model
Jiahao Yuan, Yike Xu, Jinyong Wen, Baokun Wang, Ziyi Gao, Xiaotong Lin, Yun Liu, Xing Fu, Yu Cheng, Yongchao Liu, Weiqiang Wang, Zhongle Xie
2026-02-17
Summary
This paper introduces a new way to understand users on a large scale, like for a company like Alipay, by using artificial intelligence. It focuses on creating a more dynamic and accurate 'digital fingerprint' for each user, rather than a static one, to improve how well systems can predict what users want or need.
What's the problem?
Currently, systems that try to understand users often create a single, unchanging profile for each person. This doesn't work well because people behave differently in different situations. Also, when information about users comes from many different sources – like what they buy, what they click on, and what they search for – it can be messy and conflicting, making it hard to build a clear picture of who they are. Existing methods struggle to adapt to specific tasks and handle the variety of data available.
What's the solution?
The researchers developed a system called 'Query-as-Anchor'. Instead of creating a fixed user profile, this system builds a user representation based on what the user is currently doing or asking for – the 'query'. They used powerful language models (LLMs) and a huge dataset of user behavior to train the system. They also used a clever technique called 'soft prompt tuning' to help the system focus on the most important information for each specific task. Finally, they designed the system to be fast and efficient, so it can be used in real-time applications.
Why it matters?
This research is important because it significantly improves the ability of companies to personalize experiences for their users. By understanding users in a more nuanced and dynamic way, systems can provide more relevant recommendations, better search results, and more effective services. The fact that it works well at a massive scale, like within Alipay, and doesn't slow down the system makes it practical for real-world applications and could lead to better user experiences for millions of people.
Abstract
Industrial-scale user representation learning requires balancing robust universality with acute task-sensitivity. However, existing paradigms primarily yield static, task-agnostic embeddings that struggle to reconcile the divergent requirements of downstream scenarios within unified vector spaces. Furthermore, heterogeneous multi-source data introduces inherent noise and modality conflicts, degrading representation. We propose Query-as-Anchor, a framework shifting user modeling from static encoding to dynamic, query-aware synthesis. To empower Large Language Models (LLMs) with deep user understanding, we first construct UserU, an industrial-scale pre-training dataset that aligns multi-modal behavioral sequences with user understanding semantics, and our Q-Anchor Embedding architecture integrates hierarchical coarse-to-fine encoders into dual-tower LLMs via joint contrastive-autoregressive optimization for query-aware user representation. To bridge the gap between general pre-training and specialized business logic, we further introduce Cluster-based Soft Prompt Tuning to enforce discriminative latent structures, effectively aligning model attention with scenario-specific modalities. For deployment, anchoring queries at sequence termini enables KV-cache-accelerated inference with negligible incremental latency. Evaluations on 10 Alipay industrial benchmarks show consistent SOTA performance, strong scalability, and efficient deployment. Large-scale online A/B testing in Alipay's production system across two real-world scenarios further validates its practical effectiveness. Our code is prepared for public release and will be available at: https://github.com/JhCircle/Q-Anchor.