Institute of Automation, Chinese Academy of Sciences
Incoming Ph.D. Student · Supervisor: Prof. Kaiqi Huang
Hi there Welcome to my homepage!
I'm an undergraduate student majoring in Communication Engineering at Tongji University, expecting to graduate in June 2027.
Incoming Ph.D. Student at the Institute of Automation, Chinese Academy of Sciences
Expected start: September 2027 · Supervisor: Prof. Kaiqi Huang
I explore simple, elegant ways to make AI models more capable. My interests span multimodal large language models, image/video generation, and embodied AI.
News
- Three papers submitted to ICLR 2027.
- Our work on generated image evaluation has advanced to Phase 2 of review at AAAI 2027.
- Our team received the National First Prize in the China Robot and Artificial Intelligence Competition. I served as team captain.
- Our paper, OV-OrthKD, was accepted at ACM MM 2026. I am a co-first author.
- Our paper, OrthKD, was accepted at IJCAI 2026. I am a co-first author.
- Our paper, DeepYOLO, was accepted to IJCNN 2026. I am the corresponding author.
- I received the Tongji University Outstanding Student Scholarship (Second Class) for the 2024–2025 academic year.
- Our team received the National Second Prize in the Global Campus Artificial Intelligence Algorithm Elite Competition. I served as team captain.
- Our team received the National Third Prize in the National Undergraduate Embedded Chip and System Design Competition, FPGA track. I served as team captain.
- Our team received the Provincial Third Prize in the Contemporary Undergraduate Mathematical Contest in Modeling. I served as team captain.
- I received the Provincial Second Prize in The Chinese Mathematics Competitions.
- Our team received the Provincial Third Prize in the National College Student Statistical Modeling Competition. I served as team captain.
- I received the Tongji University Outstanding Student Scholarship (Second Class) for the 2023–2024 academic year.
- I received the Provincial Second Prize in The Chinese Mathematics Competitions.
Experience
Education
Tongji University
B.Eng. in Communication Engineering
Major rank: 2/85
Research Internships
Hong Kong University of Science and Technology (Guangzhou)
Research Intern
Zhejiang University
Research Intern
Tongji University
Research Intern
Projects
Image2Other
A local-first image conversion studio with editable SVG paths and native PowerPoint exports.
JavaScriptSVGPowerPoint
What’s included
- Traces raster images into vector paths locally, with layer inspection and editable fills, opacity, and added text.
- Rebuilds diagram text and shapes through a user-configured vision model or imported Scene JSON; automatic reconstructions need checking.
- Exports SVG, native PPTX objects, and Scene JSON, alongside local PNG, JPEG, and WebP conversion with resize controls.
AutoSearch AI
Citation-aware AI search with an arXiv paper radar and an integrated PDF reader.
PythonFastAPINext.js
What’s included
- Aggregates web sources and streams answers with source citations across quick, deep, academic, and arXiv search modes.
- Connects paper discovery to summaries, method and limitation notes, and a browser-based reading workflow.
- Includes API and pipeline tests, deployment documentation, and Docker Compose configuration.
OfferPilot
An AI technical-interview practice app with follow-up questions and structured feedback.
TypeScriptNext.jsVercel AI SDK
What’s included
- Configures practice sessions by role and level, with multi-turn questioning and per-question scoring.
- Produces interview reports and stores session history locally in the current MVP.
- Supports configurable model providers, including an Ollama-based local-model option.
FPGA Algorithm Optimization
Vitis HLS algorithm optimization for the AMD FPGA competition track on a Zynq-7000 target.
C++Vitis HLSFPGA
What’s included
- Covers compression, cryptographic, and matrix-factorization kernels in the competition's Vitis Library workflow.
- Includes algorithm sources and design reports, with execution time, resource limits, and simulation correctness as engineering constraints.
Research Skills
A maintained collection of reusable workflow skills for computer science and electrical engineering research.
PythonMarkdownGitHub Actions
What’s included
- Organizes installable workflows for experiment planning, paper implementation, evaluation, and hardware development.
- Provides a machine-readable catalog, synchronization scripts, and automated checks for structure, metadata, and publication safety.
Honors and Awards
-
National First Prize
China Robot and Artificial Intelligence Competition
-
National Second Prize
Global Campus Artificial Intelligence Algorithm Elite Competition
-
National Third Prize
National Undergraduate Embedded Chip and System Design Competition
-
Provincial Third Prize
Contemporary Undergraduate Mathematical Contest in Modeling
-
Provincial Third Prize
National College Student Statistical Modeling Competition
-
Provincial Second Prize
The Chinese Mathematics Competitions
-
Second-Class Scholarship
Tongji University Outstanding Student Scholarship
Publications
* Equal contribution · Corresponding author indicated where applicable.
If You Hear It, Help Find It: Orthogonal Knowledge Distillation for Open-Vocabulary Audio-Visual Event Localization
We study open-vocabulary audio-visual event localization when visual and audio teachers differ in reliability. OV-OrthKD places their supervision in separate orthogonal feature paths and uses text prototypes for open-vocabulary transfer.
ACM MM 2026
OrthKD: Extracting Generalized Clinical Knowledge from Heterogeneous Teachers for Lightweight Deployment
We study multi-teacher distillation for diabetic retinopathy screening. OrthKD transfers complementary feature-level knowledge from heterogeneous teachers to a 5.4M-parameter MobileNetV3 student.
IJCAI 2026
Enabling Visual Reasoning in Text-Only LLMs via Symbolic Scene Descriptions
We convert images into grammar-constrained scene descriptions for text-only LLMs. On a stratified 5,000-image VQAv2 subset, DeepYOLO reaches 42.3% accuracy using 1/2.5 as many prompt tokens as the scene-graph baseline.
IJCNN 2026 Corresponding Author
From Scores to Checks: A Human-Annotated Rubric Benchmark for AI-Generated Image Faithfulness
Research on assessing whether generated images satisfy textual instructions, with a focus on reliable evaluation.
Preprint 2026
What Can Activation-Patching Evidence Establish?
Research on evaluating the evidence used to explain neural-network behavior and the reliability of the resulting interpretations.
Preprint 2026
Follow Me If You Can: Auditing Understanding-to-Generation Transfer
Research on how learning in visual understanding relates to image-generation behavior in multimodal models.
Preprint 2026
Allocation Entropy for Single-Pass Risk Triage in Neural Operators
Research on identifying unreliable predicted physical fields efficiently, with a focus on neural operators and single-pass risk assessment.
Preprint 2026
Academic Service
Reviewer
- ICLR
- PRCV
- ICME
- IJCNN
- AAAI