About
Ph.D. Student · The Chinese University of Hong Kong
I am a second-year Ph.D. student at the MoE Key Laboratory of High Confidence Software Technologies, The Chinese University of Hong Kong, supervised by Prof. Kam-Fai Wong.
My research focuses on world models, reasoning models, reinforcement learning, and multimodal learning. Before CUHK, I received my M.Eng. from Harbin Institute of Technology (Shenzhen), where I was advised by Prof. Ruifeng Xu.
News
CivWorldBench, our benchmark for evaluating language world models in Freeciv, is now available on arXiv. Explore the Project Page and code.
Our work on world model-guided reinforcement learning was accepted by EMNLP 2026 Main Conference.
I started my Ph.D. at The Chinese University of Hong Kong.
Our paper on counterfactual calibration for stance detection was accepted by NAACL 2025 Main Conference.
Our multimodal stance detection paper was accepted by ACL 2024 Findings.
Our named-entity robustness paper appeared at ICASSP 2024.
Our background knowledge-enhanced stance detection paper was accepted by EMNLP 2023 Main Conference.
Our customer-service question answering paper received the Best Paper Award at ICCC 2022.
Selected Publications
Ang Li is underlined. * denotes equal contribution.
CivWorldBench: Evaluating Language World Models in Structured Multi-Agent Sandbox Worlds
arXiv preprint · 2026
An engine-grounded benchmark in Freeciv evaluates observable forecasting, hidden-state inference, and intervention prediction, revealing gaps between forecasting strength, process correctness, and decision-making utility.
World Model-Guided Reinforcement Learning via Counterfactual User Engagement Simulation
EMNLP 2026 · Main Conference
A frozen user-engagement world model predicts counterfactual feedback and supplies dense rewards, enabling a compact 1.7B policy to match or surpass much larger LLMs without online user exposure.
A Coarse-to-Fine Text Matching Framework for Customer Service Question Answering
ICCC 2022 · Best Paper Award
A coarse-to-fine retrieval and matching pipeline improves answer accuracy while reducing inference time on customer-service datasets.
Additional Publications
Knowledge-Augmented Interpretable Network for Zero-Shot Stance Detection on Social Media
IEEE Transactions on Computational Social Systems
A Challenge Dataset and Effective Models for Conversational Stance Detection
LREC-COLING 2024