Learning to Seek Help: Dynamic Collaboration Between Small and Large Language Models
Hang Zeng, Xiangyu Liu, Yong Hu, Chaoyue Niu, Jiarui Zhang, Shaojie Tang, Fan Wu, Guihai Chen.arXiv preprint, 2026. [Download paper] [slide]
Xiangyu Liu · 刘翔宇
Senior Researcher @ WeChat AI, Tencent · Ph.D. from Peking University
About
I am a Senior Researcher at WeChat AI, Tencent. I received my Ph.D. in July 2020 from the Key Laboratory of Machine Perception and Intelligence, School of EECS, Peking University, under the supervision of Prof. Ying Tan. Before that, I earned a B.S. in Intelligence Science and Technology from Nankai University and a dual B.A. in Business Administration from Tianjin University in 2015.
My current research focuses on Large Language Models and their Applications — in particular, personalization, privacy-aware interaction, retrieval-augmented generation, and device–cloud collaborative deployment. Earlier work spans reinforcement learning, mechanism design, and evolutionary computation, with applications in e-commerce advertising and game AI.
Nothing is more practical than a good theory.
News
- 2026-09 ConvoRouter accepted to NeurIPS 2026 (Main Track, poster) — adaptive LLM routing for multi-turn conversations with continuously evolving user queries.
- 2026-04 New preprint on dynamic SLM–LLM collaboration: an on-device SLM that learns to seek help from the cloud LLM with quality, efficiency, and privacy jointly optimized (arXiv).
- 2026-04 New preprint ConvoRouter: long-horizon LLM routing for multi-turn conversations via MCTS-distilled policy learning (arXiv).
- 2026-04 Paper on automated privacy annotation in LLM interactions accepted to KDD’26 (Datasets & Benchmarks Track).
- 2025-09 RAGRouter accepted to NeurIPS 2025 — learning to route queries across retrieval-augmented LLMs.
- 2025-05 Paper on ID-free personalized LM learning accepted to KDD 2025.
📣 Hiring research interns — I am looking for motivated students to work on personal LLM agents under real interaction constraints — always-on, latency-bound, privacy-sensitive.
- Voice agents (main focus) — speech as a control interface rather than a transcription channel: telling instruction from dictation in one spoken stream, grounding disfluent colloquial commands, and multi-turn spoken revision of an evolving draft.
- Memory and continual personalization — hierarchical memory from live context to a personal knowledge base, learned from implicit behavioral signals and selectively forgettable.
- Context-grounded generation and tool use — in-situ question answering and rewriting that fuse context, retrieval, memory, and lightweight tools into the text being composed.
- Evaluating open-ended agents — data synthesis, trajectory collection, and automatic evaluation when the correct output is user-dependent.
- Device–cloud routing — local–cloud model selection under joint latency, cost, and privacy constraints (NeurIPS’25 / ‘26).
Strong background in LLMs / agents / speech / ML preferred. Remote or on-site in Beijing. Drop me an email with your CV and a short note on why you are interested.
Research Interests
- LLMs and their Applications (current focus)
- Memory agent for on-device assistants — persistent, personalized long-term memory for LLM agents (e.g., smart keyboards).
- Personalization without identifiers — learning user-adaptive models from anonymous text data. (KDD’25)
- Privacy-aware interaction — detecting and annotating privacy leakage in LLM dialogues to enable on-device protection. (KDD’26)
- LLM routing — from single-turn retrieval-augmented routing (NeurIPS’25) to long-horizon adaptive routing for multi-turn conversations (NeurIPS’26).
- Device–cloud collaborative deployment — splitting computation and knowledge between mobile devices and the cloud; on-device SLMs that learn to seek help from cloud LLMs (arXiv’26).
- Reinforcement Learning — multi-agent coordination, representation learning, exploration.
- Mechanism Design & Learning-based Auctions — for online advertising systems.
Preprints
arXiv preprint, 2026. [Download paper] [slide]
Selected Publications
Adaptive LLM Routing for Multi-Turn Conversations with Continuously Evolving User Queries
Jiarui Zhang, Xiangyu Liu, Yong Hu, Chaoyue Niu, Hang Zeng, Shaojie Tang, Fan Wu, Guihai Chen.
In NeurIPS 2026. [Download paper] [slide]Automated Annotation of Privacy Information in User Interactions with Large Language Models
Hang Zeng, Xiangyu Liu, Yong Hu, Chaoyue Niu, Fan Wu, Shaojie Tang, Guihai Chen.
In KDD 2026. [Download paper] [slide]RAGRouter: Learning to Route Queries to Multiple Retrieval-Augmented Language Models
Jiarui Zhang, Xiangyu Liu, Yong Hu, Chaoyue Niu, Fan Wu, Guihai Chen.
In NeurIPS 2025. [Download paper] [slide]Personalized Language Model Learning on Text Data Without User Identifiers
Yucheng Ding, Yangwenjian Tan, Xiangyu Liu, Chaoyue Niu, Fandong Meng, Jie Zhou, Ning Liu, Fan Wu, Guihai Chen.
In KDD 2025. [Download paper] [slide]Neural Auction: End-to-End Learning of Auction Mechanisms for E-Commerce Advertising
Xiangyu Liu, Chuan Yu, Zhilin Zhang, Zhenzhe Zheng, Yu Rong, Hongtao Lv, Da Huo, Yiqing Wang, Dagui Chen, Jian Xu, Fan Wu, Guihai Chen, Xiaoqiang Zhu.
In KDD 2021. [Download paper] [slide] [video]Optimizing Multiple Performance Metrics with Deep GSP Auctions for E-commerce Advertising
Xiangyu Liu*, Zhilin Zhang*, Zhenzhe Zheng, Chenrui Zhang, Miao Xu, Junwei Pan, Chuan Yu, Fan Wu, Jian Xu, Kun Gai.
In WSDM 2021. [Download paper] [slide] [video]Attentive relational state representation in decentralized multiagent reinforcement learning
Xiangyu Liu, Ying Tan.
In IEEE Transactions on Cybernetics (TCYB, IF=11.448) 2020. [Download paper]
A full list is available on the Publications page and on Google Scholar.
Academic Collaborations & Industry Experience
- Research collaboration, Shanghai Jiao Tong University — Device–cloud LLM applications with Dr. Chaoyue Niu. 2023 – Present
- Research collaboration, Shanghai Jiao Tong University — Learning-based mechanism design for e-commerce advertising with Dr. Zhenzhe Zheng. 2020 – 2021
- Senior Algorithm Engineer, Alibaba Group — Display Advertising — Learning-based mechanism design and auction optimization. 2020 – 2021
- Research Intern, Alibaba — Reinforcement learning in e-commerce. 2019
- Research Intern, Horizon Robotics — Reinforcement learning for autonomous vehicles and game AI. 2016 – 2019
Academic Services
- Program Committee / Conference Reviewer: WWW 2022, WCCI 2022, ECML-PKDD 2022, IJCAI 2023, KDD 2023, NeurIPS 2023, IJCAI 2024, KDD 2025, ICML 2026.
- Journal Reviewer: IEEE Transactions on Cybernetics (TCYB), Natural Computing.
Contact
- Email · liuxiangyu.george@gmail.com
- WeChat ·
shawnlue_lxy93 - GitHub · @ShawnLue