AI Researcher · CityUHK MSc AI Student
Building AI systems that are capable, efficient, and trustworthy.
Hello, I’m Chiwun (Christian) Yang (杨智桓; Yang Chi Wun in Cantonese). I study the mechanisms and theory of modern machine learning, with a focus on large language models, reasoning and agents, efficient training and inference, and trustworthy AI.
I am an MSc student in Artificial Intelligence at City University of Hong Kong and a Research Assistant at the Hong Kong Institute of AI for Science, working with Prof. Ning Miao on AI for mathematics, LLM reasoning, and agentic reinforcement learning. I received my BEng in Artificial Intelligence from Sun Yat-sen University in 2025.
Research
My research combines theory and experiments to understand how modern machine learning systems learn, reason, and use information. I am interested in LLM inference, fine-tuning, reasoning and agents, the science of deep learning, and trustworthy and data-private machine learning. I especially enjoy connecting mathematical structure to measurable model behavior and working implementations, and I am open to new directions such as world models.
01
Efficient LLMs & Agents
Sparse attention, KV-cache compression, speculative decoding, parameter-efficient fine-tuning, and systems for reasoning and agents.
02
Trustworthy & Secure AI
Copyright-aware optimization, data privacy and recovery, calibration, and the foundations of safe and reliable model behavior.
03
Learning Dynamics & Theory
Scaling laws, optimization and generalization, representation geometry, and how pretraining and reward learning shape model capabilities.
Publications & Preprints
* denotes equal contribution in alphabetical order where marked. For the complete and most current record, see Google Scholar.
- AAAI 2024 Timothy Chu*, Zhao Song*, and Chiwun Yang*. How to Protect Copyright Data in Optimization of Large Language Models? Code
- EMNLP 2025 · ICLR 2025 SCOPE Workshop Oral Yingyu Liang*, Zhenmei Shi*, Zhao Song*, and Chiwun Yang*. Towards Infinite-Long Prefix in Transformer. Code
- NeurIPS 2025 Yang Cao*, Xiaoyu Li*, Zhao Song*, and Chiwun Yang*. Efficient k-Sparse Band-Limited Interpolation with Improved Approximation Ratio.
- ICML 2025 Jing Xiong, Jianghan Shen, Chuanyang Zheng, Zhongwei Wan, Chenyang Zhao, Chiwun Yang, Fanghua Ye, Hongxia Yang, Lingpeng Kong, and Ngai Wong. ParallelComp: Parallel Long-Context Compressor for Length Extrapolation. Code
- CPAL 2025 Majid Daliri*, Zhao Song*, and Chiwun Yang*. Unlock the Theory behind Scaling 1-bit Neural Networks.
- CPAL 2025 Yekun Ke*, Yingyu Liang*, Zhenmei Shi*, Zhao Song*, and Chiwun Yang*. Curse of Attention: A Kernel-Based Perspective for Why Transformers Fail to Generalize on Time Series Forecasting and Beyond.
- ICLR 2025 SLLM Workshop Yichuan Deng*, Zhao Song*, Jing Xiong*, and Chiwun Yang*. How Sparse Attention Approximates Exact Attention? Your Attention is Naturally nC-Sparse.
- ICLR 2025 DeLTa Workshop Yang Cao*, Zhao Song*, and Chiwun Yang*. Video Latent Flow Matching: Optimal Polynomial Projections for Video Interpolation and Extrapolation.
- Preprint · 2026 Chiwun Yang and Xiaoyu Li. What Pretraining and Midtraining Make Learnable from Rewards?
- Preprint · 2026 Chiwun Yang and Xiaoyu Li. One Spectrum, Two Resources: Data-Memory Scaling in Autoregressive Prediction
- Preprint · 2026 Chiwun Yang and Xiaoyu Li. Beyond Sparse Weights: When Is Attention Compressible?
- Preprint · 2026 Xiaoyu Li and Chiwun Yang. Recalibrating Many Questions with Few Labels: Minimax Rates and Identification Limits
- Preprint · 2026 Xiaoyu Li, Zhizhou Sha, Chiwun Yang, and Dai Shi. A Categorical Atlas of Parameter-Efficient Fine-Tuning: Cones, Orbits, and Gauges
- Preprint · 2026 Xiaoyu Li, Zhizhou Sha, and Chiwun Yang. The Birkhoff Geometry of Manifold-Constrained Hyper-Connections: Two Channels, Vertex Viscosity, and Sinkhorn as a Retraction
- Preprint · 2023 Yichuan Deng*, Zhao Song*, Shenghao Xie*, and Chiwun Yang*. Unmasking Transformers: A Theoretical Approach to Data Recovery via Attention Weights.
- Preprint · 2023 Chenyang Li*, Zhao Song*, Weixin Wang*, and Chiwun Yang*. A Theoretical Insight into Attack and Defense of Gradient Leakage in Transformer.
- Preprint · 2025 Jiangxuan Long*, Zhao Song*, and Chiwun Yang*. Theoretical Foundation of Flow-Based Time Series Generation: Provable Approximation, Generalization, and Efficiency.
- Preprint · 2025 Chiwun Yang. Unifying Learning Dynamics and Generalization in Transformers Scaling Law.
Open Source
LOLPingYourAI
A desktop AI teammate that turns a ping into a screenshot, window context, and task intent for Codex, Claude, or DSH. Built with Electron and React, with persistent sessions, interruption controls, follow-ups, and recovery across app restarts.
LightCone-Spec
A framework for online drafter adaptation in speculative decoding. Uses target-model feedback to adapt the drafter at request boundaries, with an SGLang-based runtime, exactness checks, telemetry, and resumable experiments.
Other projects include DLT Proof Writing Skill for structured mathematical drafting and review, and fast_yolov7_pytorch for pruning and quantization. Research implementations are linked alongside the papers above.
Honors & Awards
Selected scholarships, reviewer recognition, and competition awards.
ICML 2026 Silver Reviewer
Recognition for reviewing contributions to the International Conference on Machine Learning.
Hong Kong Future Talents Scholarship Scheme for Advanced Studies (FTSS)
Merit-based HKSAR Government scholarship for postgraduate study in priority areas.
Guangdong Provincial Third Prize, CUMCM
Provincial recognition in the Contemporary Undergraduate Mathematical Contest in Modeling.
Silver Award, “Win at SYSU” Innovation and Entrepreneurship Competition
University-wide competition evaluating technical innovation, social value, and application potential.
Second Prize, Alibaba Lingxi Games Programming Competition
Industry-sponsored recognition for programming and technical problem solving.
Second Prize, SYSU AI Challenge
Sun Yat-sen University artificial intelligence competition.
Education & Experience
Education
Research
Jun – Dec 2025
Engineering
Academic Service
Reviewer for NeurIPS 2026 (main conference and workshop), AAAI (2026–2027), ICLR (2025–2026), ICML (2025–2026; ICML 2026 Silver Reviewer), COLM (2025–2026), and AISTATS 2026, with additional reviewing service for ICLR 2025 workshops and a CVPR 2026 workshop.
Let’s talk
Ideas, criticism, and collaborations are welcome.
I am always happy to discuss research on efficient and trustworthy AI, especially work that connects theory with systems.
christiannyang37 [at] gmail [dot] comLast updated: October 25, 2026.