Yihui Wang

Yihui Wang

Also called Mistry-Nihilityn. Research Intern @ NUS NExT++.

Yihui Wang’s Homepage

👋 About

I am an undergraduate student in Software Engineering at Hefei University of Technology, and a remote research intern at NExT++ Research Centre, National University of Singapore, supervised by Dr. Yonghui Yang and Prof. Tat-Seng Chua (NUS Chair Professor, Singapore National Academy of Sciences). My research focuses on LLM Safety & Alignment, Interpretable Robust Learning, and Multimodal Deepfake Detection. I also hold multiple ICPC/CCPC/CCSP medals, including a CCPC National Gold, which sharpen my algorithmic and engineering skills.

📰 News

  • [May. 2026] Led HFUT Team 101 to win a Gold Medal at the CCPC National Invitational (Nanchang).
  • [Jun. 2026] Won a Silver Medal at the CCPC National Invitational (Guiyang) as a member of Exercise Everyday.
  • [May. 2026] Our work NeVA (Neuron-level Value Alignment) was submitted to ACL Rolling Review (ARR) (CCF-A, co-first author).
  • [Apr. 2026] Our work S³ (Suppressing Method-Specific Shortcuts) was submitted to ACM MM 2026 (CCF-A, first author).
  • [Feb. 2026] Our work CLEAR (Cross-modal De-redundancy) has been transferred to AAAI (CCF-A).
  • [2025] Ongoing research on LLM Deep Safety Alignment (BEACON), targeting ICLR 2027 (first author).

🎓 Education

  • B.Eng. in Software Engineering, Hefei University of Technology, 2023 – 2027
    GPA: 90.78/100 (Top 1.75%)
  • Remote Research Intern, NExT++ Research Centre, National University of Singapore, 2025 – Present
    Supervised by Dr. Yonghui Yang and Prof. Tat-Seng Chua

🔬 Research

All work conducted at NUS NExT++ Research Centre.

BEACON Framework Overview

LLM Deep Safety Alignment: Auditing Shallow Alignment via Representation-Level Diagnosis (BEACON)

First Author · Targeting ICLR 2027 (CCF-A) · 2025 – Present

  • Quantified a belief-behavior gap: internal probes detect harmful intent with >95% accuracy at any prefix length, yet model compliance surges from <2% to >50% as prefix grows — proving alignment failure is a generation-side problem, not a representation-side problem.
  • Proposed BEACON: token-wise consistency losses anchor the model’s generation behavior to its own internal harmfulness judgment, maintaining refusal robustness regardless of prefix length. Cuts adversarial attack success rate from >50% to ≤5% across four models and up to 3,000-token prefixes.

S³ Method Overview

S³: Suppressing Method-Specific Shortcuts for Generalizable Deepfake Detection

First Author · ACM MM 2026 (CCF-A, Under Review) · 2025 – 2026

  • Identified the root cause of cross-domain failure in deepfake detection: trained detectors learn to classify forgery method rather than forgery presence — a shortcut that makes out-of-distribution collapse structurally inevitable, regardless of backbone or data scale.
  • Designed two independent remedies: gradient projection (NSP) steers training away from method-discriminative feature directions; a plug-in inference module (NAE) achieves the same effect without backbone retraining. Cross-domain AUC gains exceed 10× the in-domain cost, validated across four backbones and a fully held-out open-world benchmark.

NeVA Method Overview

NeVA: Controllable Value Alignment in LLMs via Neuron-Level Editing

Co-first Author · ACL ARR (CCF-A, Under Review) · 2026

  • Locates and edits the sparse subset of neurons governing specific value outputs in LLMs and LVLMs, enabling fine-grained value alignment without parameter updates.

CLEAR: Null-Space Projection for Cross-Modal De-Redundancy in Multimodal Recommendation

Co-first Author · AAAI (CCF-A, Submitting) · 2025 – 2026

  • We remove coarse-grained cross-modal redundancy by projecting embeddings onto the null-space of the dominant covariance subspace — plug-and-play, no auxiliary loss.

🏆 Awards & Honors

  • National Scholarship, Top 1% — Ministry of Education of China (2024)
  • Gold Medal — CCPC National Invitational, Nanchang (2026)
  • Silver Medal — CCPC National Invitational, Guiyang (2026)
  • Bronze Medal — ICPC Asia Regional, Chengdu (2025)
  • Bronze Medal, 2× — CCPC Regional, Jinan & Zhengzhou (2025)
  • Bronze Medal — CCF CCSP; CSP 360/500, Top 1% (2025)
  • First Prize, 2× — Anhui Province Robot Competition, Programming Track (2024, 2025)
  • Gold Prize — Anhui Province Innovation Competition (2024)

👥 Leadership & Service

  • ACM Team Leader, HFUT XCPC Team (Binary Algorithm Studio), 2024 – Present
    CCPC National Gold Medalist; lead daily training & solution-sharing for 20+ members; problem setter and organizer for HFUT ACM School Contest and Anhui Province Programming Competition (100+ teams).
  • Video Team Lead, HFUT University Media Center — Visual Imaging Studio, 2023 – 2025
    Produced 10+ official university videos, several adopted by Xinhua News Agency; organized cross-college photography and film editing courses reaching 300+ students.

💻 Technical Strengths

  • Research Focus: LLM Safety & Alignment · Robust Deepfake Detection · Interpretability
  • Engineering: CCPC National Gold Medalist; production-grade full-stack developer in Python, C/C++, Java, Qt, and modern web frameworks; ship complex systems from prototype to deployment; 2 software copyrights
  • Tools & Languages: PyTorch · MATLAB · LaTeX · Git · English CET-6: 553
  • Open Source & Community: Maintainer of ManimVE rendering engine; educational math/CS animation creator on Bilibili; active open-source contributor

📫 Contact