Zeteng Lin
Portrait of Zeteng Lin

Zeteng Lin

Ph.D. Candidate, Data Science and Analytics

Information Hub, The Hong Kong University of Science and Technology (Guangzhou)

I am a Ph.D. candidate in Data Science and Analytics at HKUST(GZ), supervised by Prof. Jing Tang and Meta Scientist Yujun Cai. My Ph.D. committee chair is Prof. Danyang Xie and Prof. Tang Nan.

My research focuses on trustworthy multimodal AI, particularly on evaluating what foundation models truly perceive, infer, and simulate across images and videos.

Experience

  • Co-founder & Founding Engineer, CashFlowApp, 2026-present. Building an LLM-powered strategy-authoring and validation service for AI-assisted portfolio workflows, including natural-language-to-config generation, structured-output validation, self-healing validation loops, risk-field gating, numeric-claim checking, grounded chat, and backtest explanation.
  • Member, Macau International Artificial Intelligence Association (MIAIA; 澳门国际人工智能学会).
  • M.Phil. degree supervised by Prof. Jing Tang and Prof. Yangqiu Song.
Earlier Research and Service private

Selected Publications

("#" indicates co-first authors, and "*" indicates the corresponding author.)

  1. Emission-GPT: A domain-specific language model agent for knowledge retrieval, emission inventory and data analysis
    Jiashu Ye, Tong Wu, Weiwen Chen, Hao Zhang, Zeteng Lin, Xingxing Li, Shujuan Weng, Manni Zhu, Xin Yuan, Xinlong Hong, Jingjie Li, Junyu Zheng*, Zhijiong Huang, and Jing Tang.
    arXiv:2510.02359, 2025. arXiv
    PaperCode
  2. Unspoken Details: Inferring Hidden Causality and Retrieving Domain-Specific Knowledge for Image Generation
    Wen You, Zhijun Ma, Zeteng Lin, and Troy Tianyu Lin*.
    Proceedings of the 18th International Symposium on Visual Information Communication and Interaction (VINCI), 2025. DOI: 10.1145/3769534.3769595. ACM VINCI 2025
    PaperCode

Working Papers

  1. Causal-OPD: Counterfactual Branch Credit Assignment for On-Policy Distillation
    Selects teacher decisions for on-policy distillation by scoring matched student and teacher continuations from the same prefix with a task verifier.
    Xihui Wu, Bolin Zeng, and Zeteng Lin.
    Working paper, 2026. Working Paper
    PaperCode
  2. ReCam-VL: Reversal-Aware VLMs for Camera Motion Understanding
    Tests whether VLMs understand camera-motion direction by evaluating paired forward and reversed videos under a reversal-closed motion taxonomy.
    Zeteng Lin, Xihui Wu, Hao Li, Yujun Cai, and Jing Tang.
    Under review at AAAI 2027. Under Review
    PaperCode
  3. MoTIE: Motion Tool-Augmented Inference with Evidence for Camera-Motion Understanding
    Provides frozen VLMs with explicit optical-flow, frame-difference, drift, scale, and parallax evidence through a structured Motion Blueprint.
    Zeteng Lin, Xihui Wu, Yujun Cai, and Jing Tang.
    Working paper, 2026. Working Paper
    PaperCode
  4. VUG-RAG: Efficient and Fine-Grained Videos RAG using Video Understanding Graph
    Builds a hierarchical Video Understanding Graph for long-video retrieval, preserving global descriptions, visual attributes, object features, and relations.
    Zeteng Lin, Xihui Wu, Junhao Li, Yujun Cai, and Jing Tang.
    Working paper, 2026. Working Paper
    PaperCode
Computational Social Science 3 publications, including SSCI-Q1
  1. Understanding Generational Shifts and Democratic Support in China: Value Change vs. Life Satisfaction
    Jie Chen, Narisong Huhe*, Zeteng Lin.
    Journal of Contemporary China, 2025. DOI: 10.1080/10670564.2025.2597751. SSCI-Q1
    PaperCode
  2. 中国民众权利义务认知的代际差异: 基于价值观与生活满意度的中介效应分析
    Chen Jie, Narisong Huhe#, and Zeteng Lin*.
    Journal of Modernization Studies, 4(1): 78-92, 2025. DOI: 10.20200/j.cnki.gjxdh.2025.01.003. CSSCI扩展版
    PaperCode
  3. 大语言模型能否成为可靠的硅基受访者？——基于 CFPS 回答分布的微调训练实验
    Can Large Language Models Become Reliable Silicon Respondents? A Fine-tuning Experiment Based on CFPS Response Distributions
    Lin Zeteng, Ye Zhenye, and Wang Renhe.
    Journal of Global ESG Innovation, Vol. 1, Issue 1, p. 5, 2026. Published
    PaperCode

Open Source

Desensitized engineering prototypes and research systems. Projects marked Ascend NPU were verified on Ascend 910C hardware, not merely documented as compatible.

  • NexusFlow assistant workflow studio

    A visual workflow studio for assistant pipelines — intent routing, retrieval, tool calls and model orchestration — running on self-hosted vLLM on Ascend 910C.

    Ascend NPU Workflow TypeScript On-device AI
  • visual-document-rag layout-aware document RAG

    RAG for visually rich documents where layout, formulas and tables survive the pipeline, ported from CUDA to Ascend NPUs and tested on 910C.

    Ascend NPU RAG Document AI CANN
  • AgenticVideoRAG long-video reasoning

    An agentic framework for long-video question answering that plans missing evidence, retrieves targeted clips, and verifies grounding before answering.

    Video QA Agents RAG
  • Factor-RAG structure-aware chunking

    Structure-aware Markdown chunking that keeps tables, headers and figure captions intact, with a reproducible benchmark scoring any chunking strategy on structural fidelity.

    RAG Document AI Python
  • SoloChart conversational analytics

    A single-file page that turns a plain-language question into an interactive chart, backed by self-hosted vLLM on Ascend 910C and carrying no API credential.

    Ascend NPU Analytics vLLM Zero-build
  • Cantonese ASR speech recognition

    A reproducible Cantonese speech-recognition system fine-tuned and distilled from Whisper-small, from data preparation through offline packaging.

    Speech Whisper Low-resource
  • Silicon Respondents on CFPS survey simulation

    Fine-tunes language models against real CFPS response distributions so generated survey answers match a population's option proportions, not just its modal answer.

    LoRA Computational Social Science Qwen
Books and Chapters 7 entries
  • 高洪波, 林泽腾. (2025). 2024年传媒业投融资发展报告[M]// 杨涛, 金巍 编. 中国文化金融发展报告（2024-2025）. 北京: 社会科学文献出版社.
  • 陈能军, 林泽腾, 毛君婕. (2025). 人工智能如何赋能智慧医疗？——基于全国多城市的案例分析[M]// 中国（杭州）智慧城市研究院. 中国智慧城市发展蓝皮书（2025）. 杭州: 浙江大学出版社.
  • 陈能军, 林泽腾. (2025). 国内大循环中的文化消费新动能[M]//广东消费经济学会2024年论坛蓝皮书.
  • 陈能军, 林泽腾, 戎涛. (2024). 智慧城市“城设”赋能城市文明典范建设[M]//李凤亮 编. 全球城市文明发展报告(2023): 数字时代的全球城市文明. 社会科学文献出版社. ISBN: 978-7-5228-3727-7.
  • 陈能军, 林泽腾, 高洪波. (2024). 深圳应抢占大模型备案关键窗口，避免人工智能创业资源流失[N]. 深圳社科智库专报, 2024(118).
  • 陈捷, 林泽腾. (2022). 以全国文明典范城市创建为深圳高质量发展赋能[N]. 深圳特区报, 2022-09-06(B01). DOI: 10.28776/n.cnki.nsztq.2022.004194.
  • 陈刚, 项后军, 林泽腾 等. (2018). 广东省地方金融监管发展报告(2018)[M]. 广东经济出版社.

Teaching

  • Teaching Assistant Lead, AI Startup Group, Greater Bay Area Youth Internship Program, HKUST(GZ) CMA Lab, Summer 2026.
  • Teaching Assistant, DSAA6000Q Large Language Models: Theory and Practice, HKUST(GZ), Spring 2026.
  • Teaching Assistant, HKUST(GZ) Red Bird Challenge Camp, Fall 2024.
Earlier Teaching private
  • Teaching Assistant, Innovation, Policy and Entrepreneurship, HKUST(GZ), Spring 2023.
  • Teaching Assistant, Introduction to International Relations Theory, SUSTech, Spring 2022.
  • Teaching Assistant, Foundation of Comparative Political Science, SUSTech, Spring 2022.

Others

Selected Talks & Posters

  • Selected Finalist, 2nd Greater Bay Area Joint Computer Science Graduate Research Poster Competition, Shenzhen University, 2026. Poster: "MoTIE: Motion Tool-Augmented Inference with Evidence for Camera-Motion Understanding."
Earlier Social Science Talks 4 talks
  • Political Psychology in East Asia Workshop, Chinese University of Hong Kong, 2024. Presented "What Chinese people really want politically: procedures and substances."
  • APSA Chinese Politics Mini-Conference, University of Pennsylvania, 2024. Presented "What Chinese people really want politically: procedures and substances."
  • PSA Annual International Conference, 2024. Presented "What Chinese people really want politically: procedures and substances."
  • National Computational Social Science Forum, 2023. Presented work on computational social science and public-sector AI applications.

Honors

  • HKUST Postgraduate Scholarship.
  • GAC Group-HKUST(GZ) Innovation Talent Development Scholarship, 2024.
  • National Scholarship.
  • MCM Honorable Mention.
  • Shenzhen Kunpeng Young Innovative Talent.
Media Coverage 3 items

Selected Patents

  • Zeteng Lin et al. 大语言模型推理优化方法、装置、终端设备及存储介质 / Method, apparatus, terminal device, and storage medium for optimizing large language model inference. Chinese invention patent, granted 2026.
  • Zeteng Lin et al. 一种边缘设备语音识别的动态能耗优化方法、装置及介质 / Dynamic energy consumption optimization method, apparatus, and medium for speech recognition on edge devices. Chinese invention patent, granted 2026.
  • Zeteng Lin et al. 一种适用于多语种多语音的跨模态转换方法、系统及介质 / Cross-modal conversion method and system suitable for multiple languages and multiple voices, and medium. Chinese invention patent, granted 2026.
  • Zeteng Lin et al. 一种解决大型语言模型幻觉问题的方法和系统 / Method and system for solving hallucination problems in large language models. Chinese patent application, 2025.
  • Zeteng Lin et al. 智能陪伴机器人系列（智能陪伴机器人；桌面机器人“萌萌心语”；智能陪伴机器人主机；家居陪伴机器人） / Intelligent companion robot series (intelligent companion robot; desktop robot "Mengmeng Xinyu"; companion robot host; home companion robot). Chinese design patents, granted 2025.