Ph.D. Student · Harbin Institute of Technology, Shenzhen
I study how large language models represent, reason about, and align with human cognitive states in social contexts.
I am a Ph.D. student in the School of Computer Science and Technology and the Research Institute of Cyberspace Security at Harbin Institute of Technology, Shenzhen, advised by Prof. Qing Liao. My research connects LLMs, agents, and cognitive modeling, with a focus on social behavior prediction, cognitive-state modeling, emotional support dialogue, and privacy-preserving recommendation.
我关注大语言模型的底层表征与认知建模：模型如何编码人的情感、立场、思维风格、意图等认知状态，如何在 情感支持对话中进行社会认知推理，以及如何让模型更稳健地服务真实社交场景中的理解、推荐与交互任务。
Research Agenda
My long-term goal is to build language agents that do not merely imitate surface text, but maintain useful internal representations of human cognitive states and use them for socially grounded reasoning and interaction.
Model emotions, stances, intentions, thinking styles, and other socially meaningful cognitive signals.
Use structured reasoning and multi-agent interaction to support social understanding and prediction.
Apply cognitive modeling to emotional support dialogue, recommendation, and social behavior prediction.
Selected Publications
A selected list is shown here; please visit Google Scholar for the complete and most recent publication list.
Proposes a cognitively grounded benchmark and modeling framework for analyzing how LLMs represent multiple simultaneous cognitive states, especially when social signals crowd or interfere with each other.
Constructs CogBiasESC and proposes CoPoLLM to diagnose cognitive distortions in help-seeker utterances and guide safer, deeper cognitive interventions in emotional support dialogue.
Introduces a collaborative multi-agent framework that converts trajectory signals into language, narrows the candidate POI space, and improves next-location prediction with specialized agents.
Studies cross-domain sequential recommendation in privacy-preserving federated settings, connecting multi-platform user behavior modeling with robust recommendation transfer.
Uses hypergraph signal decoupling for federated cross-domain recommendation, improving how heterogeneous user-item signals are transferred across domains without centralizing private data.
Explores prompt-guided distillation for personalized federated learning under model heterogeneity, aiming to improve cross-client knowledge transfer without forcing a single model architecture.
Published Themes
Studies how LLMs represent emotion, stance, thinking style, and intention, and how representation spaces affect joint cognitive-state understanding.
Builds cognitive-policy-driven LLMs for emotional support dialogue, with attention to distortion diagnosis, intervention strategy, and safety-aware response generation.
Designs specialized agents that bridge trajectory data and language reasoning for user mobility and next-location prediction.
Develops federated recommendation and personalized learning methods for cross-domain, multi-platform, and heterogeneous-client settings.
Contact
I am especially interested in collaborations around LLM representations, cognitive modeling, socially grounded agents, emotional support dialogue, and privacy-preserving learning. I am also building a long-term academic profile toward future faculty-track opportunities.