Hi, I'm Yichi — I study when multimodal and embodied models break, and how to measure it.

I am an undergraduate in Artificial Intelligence at Beijing University of Technology (Project 211 & Double First-Class), and a research intern at HKUST (Guangzhou) advised by Prof. Renjing Xu. My work sits where trustworthy multimodal AI meets embodied intelligence — probing how vision-, language-, and action-models fail under adversarial and physical stress, and building benchmarks that make those failures measurable. So far I've contributed to four published / accepted papers and two works under review, including a Best Student Paper Award at the IJCAI 2025 workshop.

1. Jan 2026 Began a research internship at HKUST (Guangzhou) with Prof. Renjing Xu. → Research
2. 2026 Two papers under review at NeurIPS 2026 D&B and ACM MM 2026. → Publications
3. 2026 Won a Bronze Medal (solo) at Kaggle BirdCLEF+ 2026 — 299 / 4084 teams. → Honors
4. Sep 2025 Jailbreak-AudioBench accepted at NeurIPS 2025 Datasets & Benchmarks Track. → Paper
5. Aug 2025 Received the Best Student Paper Award at the IJCAI 2025 workshop for TVPI. → Paper
6. 2025 Transfer Attack for Bad and Good accepted at ACM Multimedia 2025. → Paper
7. 2025 Led my team to 5th / 258 (Top 2%) in the IJCAI 2025 Workshop Challenge. → Honors

• ◆Contributed to four published / accepted works and two works under review (NeurIPS 2026 D&B; ACM MM 2026).
• ◆Focus on the adversarial robustness of multimodal LLMs and the reliability of embodied world models — from typographic prompt injection to whole-body benchmarking.

Open to PhD opportunities & research collaboration.