cv

MS

Stanford University

Computer Science (GPA: 4.0/4.0)

BE

Tsinghua University

Industrial Engineering (CS GPA: 3.8/4.0)

Senior Machine Learning Engineer

Apple

Working on LLM post-training and RL alignment for large-scale retrieval. Focused on reward design, policy optimization, and preference-data workflows that move beyond traditional ranking toward reasoning-driven candidate generation.

• LLM post-training (SFT/RL)
• Reward design & RL policy optimization
• Generative retrieval
• Alignment evaluation

Applied Scientist

Microsoft AI

Worked on LLM evaluation, RAG alignment, and large-scale multi-modal retrieval. Adapted CLIP, SigLIP, and Perception Encoder to multiple text-to-image and image-to-image retrieval/reranking tasks; trained teacher models and distilled into lightweight students for production. Designed GPT-based RAG evaluation and human-in-the-loop alignment infrastructure for LLM Copilot systems.

• LLM evaluation & RAG alignment
• Large-scale multi-modal retrieval (CLIP/SigLIP)
• Knowledge distillation
• Human-in-the-loop evaluation

Data & Applied Scientist Intern

Microsoft AI

Worked on large-scale multi-modal Transformer compression. Extended structured pruning methods to CLIP (Contrastive Language-Image Pre-training) for the first time, reducing model size by 40% with minimal accuracy loss; built sparse PyTorch training pipelines with recovery schedules and auxiliary losses to stabilize multi-modal compression.

• Multi-modal Transformers
• Large-scale model compression
• Structured pruning on CLIP

Machine Learning Engineer

ByteDance AI Lab (later Seed)

Worked on NLP and data-centric safety at AI Lab (later Seed), supervised by Dr. Hang Li. Built large-scale fake-news and harmful-content detection pipelines with finetuned BERT/NLI models, downstream human review, and 2M+ guideline-driven annotations; deployed with auto-retraining and version-controlled rollout.

• NLP & data-centric safety
• Large-scale harmful-content detection
• BERT/NLI fine-tuning
• Guideline-driven annotation

ColBERT-Serve: Efficient Multi-stage Memory-Mapped Scoring

Advances in Information Retrieval: 47th European Conference on Information Retrieval, ECIR 2025

We study serving retrieval models, particularly late interaction retrievers like ColBERT, to many concurrent users at once and under a small budget, in which the index may not fit in memory. We present ColBERT-serve, a serving system that applies a memory-mapping strategy to the ColBERT index, reducing RAM usage by 90% and permitting its deployment on cheap servers, and incorporates a multi-stage architecture with hybrid scoring, reducing ColBERT's query latency and supporting many concurrent queries in parallel.

DeepThink: Aligning Language Models with Domain-Specific User Intents

arXiv preprint arXiv:2502.05497

We propose DeepThink, a framework for aligning large language models with domain-specific user intents through targeted instruction synthesis and preference optimization, improving instruction-following quality on specialized tasks.

Overview of the Ninth Dialog System Technology Challenge: DSTC9

IEEE/ACM Transactions on Audio, Speech, and Language Processing

Overview paper for the Ninth Dialog System Technology Challenge (DSTC9), covering multi-domain task-oriented dialogue systems and cross-lingual dialogue state tracking.

Investigating Effect of Dialogue History in Multilingual Task-Oriented Dialogue Systems

arXiv preprint arXiv:2112.12318

We study how dialogue history influences multilingual task-oriented dialogue systems, analyzing cross-lingual transfer and the role of context length on end-to-end dialogue performance.

Multi-domain Task-oriented Dialog Challenge II at DSTC9

AAAI-2021 Dialog System Technology Challenge 9 Workshop

The paper provides an overview of the 'Multi-Domain Task Completion Dialog Challenge II' track at the 9th Dialog System Technology Challenge (DSTC9). Two tasks are introduced in this track: end-to-end multi-domain task completion and cross-lingual dialog state tracking.

A Large-Scale Chinese Short-Text Conversation Dataset

Natural Language Processing and Chinese Computing (NLPCC)

We present a large-scale cleaned Chinese conversation dataset LCCC, which contains a base version (6.8 million dialogues) and a large version (12.0 million dialogues). The quality of our dataset is ensured by a rigorous data cleaning pipeline, which is built based on a set of rules and a classifier that is trained on manually annotated 110K dialogue pairs.

KdConv: A Chinese Multi-domain Dialogue Dataset Towards Multi-turn Knowledge-driven Conversation

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL)

We propose a Chinese multi-domain knowledge-driven conversation dataset, KdConv, which grounds the topics in multi-turn conversations to knowledge graphs. Our corpus contains 4.5K conversations from three domains (film, music, and travel), and 86K utterances with an average turn number of 19.0.

CrossWOZ: A Large-Scale Chinese Cross-Domain Task-Oriented Dialogue Dataset

Transactions of the Association for Computational Linguistics (TACL)

To advance multi-domain (cross-domain) dialogue modeling as well as alleviate the shortage of Chinese task-oriented datasets, we propose CrossWOZ, the first large-scale Chinese Cross-Domain Wizard-of-Oz task-oriented dataset. It contains 6K dialogue sessions and 102K utterances for 5 domains, including hotel, restaurant, attraction, metro, and taxi.

Stanford University

Collaborated with Prof. Christopher Potts, Prof. Matei Zaharia, and Prof. Kwabena Boahen on efficient large-scale neural retrieval. Designed memory-mapped multi-stage scoring for ColBERT, reducing RAM usage by 90% while preserving retrieval quality. Published at ECIR 2025.

• Large-scale neural retrieval (ColBERT)
• Memory-mapped multi-stage scoring
• ECIR 2025

Stanford University

Collaborated with Prof. Tengyu Ma on model-based RL for task-oriented dialogue. Built SLBO and Vanilla Policy Gradient (VPG) agents with user-simulator training pipelines for dialogue policy optimization. Part of Stanford UGVR program.

• Reinforcement learning (SLBO/VPG)
• Dialogue policy optimization
• Model-based RL

Tsinghua University

Collaborated with Prof. Minlie Huang on large-scale Chinese dialogue datasets and multi-turn knowledge-driven conversation modeling. Co-authored multiple papers at ACL, TACL, and NLPCC.

• Large-scale dialogue datasets
• Multi-turn knowledge-driven conversation
• ACL/TACL/NLPCC publications

Reviewer

Conducted 45+ paper review work for top-tier conferences and journals in natural language processing (NLP) and computer vision (CV), including EMNLP'23, ICLR'24, WACV'24, COLING'24, SIGIR'24, KDD”24, CIKM'24, WACV'25, ICLR'25, COLING'25, ACL'25, Computer Speech & Language.

• Academic Service

NLPCC Best Student Paper

Natural Language Processing and Chinese Computing (NLPCC)

NLPCC is a leading international conference specialized in the fields of Natural Language Processing (NLP) and Chinese Computing (CC).

Stanford UGVR Scholar

Stanford University, Tsinghua University

Highly selective summer research fellowship at Stanford Engineering, admitting up to 18 top undergraduates from Tsinghua, Peking, and other leading Chinese universities each year. Awardees are matched with Stanford faculty to conduct full-time research.

1st Prize in National Olympiad in Informatics in Provinces

National Olympiad in Informatics

The National Olympiad in Informatics in Provinces (NOIP) is a competitive programming competition for high school students in China. It is one of the most prestigious and challenging competitions in the field of computer science.