Yuxiang Wang

Yuxiang Wang

MS @ Fudan University. Research interests: LLMs (Agents), Diffusion Language Models, AI4S, Multimodal LLMs, GNNs.

Yuxiang Wang

I am a first-year Master’s student at the School of Data Science, Fudan University, where I am very fortunate to be advised by Prof. Baojian Zhou at the Knowledge Works Lab.

I received my B.Eng. in Computer Science (Honors Class) from the Zhuoyue Honor College, Hangzhou Dianzi University in June 2025, where I worked with Prof. Wang.

Research Interests

My research interests span:

  • LLMs (Agents) — post-training, tool use, multi-turn reasoning (Thought→Action→Observation), and RL for agents
  • Diffusion Language Models — locality-aware architectures, training stability, and the inductive-bias / task-structure relationship
  • AI4S — protein/small-molecule modeling, unified evaluation frameworks, and Protein Agents
  • Multimodal LLMs — unified understanding & generation, e.g. protein ↔ text bidirectional modeling
  • Graph Neural Networks — dynamic-graph reasoning, fraud detection on spatio-temporal graphs

News

  • 2026 — Sole first author on Locality-aware Diffusion Language Modeling (introducing the Scatter and Jigsaw blockwise architectures); second author on Semantic Diffusion Language Modeling — both studying training stability and architecture design of diffusion language models.
  • 2025.10 — Joined Alibaba (AI4S / Multimodal LLM) as a research intern.
  • 2025.09 — Started my MS at Fudan University, Knowledge Works Lab.
  • 2025.06 — Graduated from HDU as an Outstanding Graduate; admitted to Fudan University via recommendation.

Selected Publications

  • Locality-aware Diffusion Language Modeling — Sole first author. Studies the trainability of Masked Diffusion LMs and proposes two locality-aware blockwise architectures (Scatter and Jigsaw) bridging AR and Diffusion regimes.
  • Spatio-Temporal Distance and Frame-Based Dynamic Graph Fraud Detection (STM) — First author, CCF-B venue. Frame-wise intra/inter aggregation over dynamic graphs achieves SOTA on multiple benchmarks. Resulted in an invention patent.
  • Semantic Diffusion Language Modeling (SemDLM) — Second author. Semantic-neighborhood diffusion + shared refresh branch reaches 27.19 Test PPL on LM1B.

You can find my full Curriculum Vitae for more details.

Open-Source Study Notes — Welcome to Learn!

I have open-sourced my study notes covering LLM, Multimodal, RL alignment, and NLP. Everyone is welcome to use them for learning and interview preparation!

NotesDescriptionLink
LLM & Multimodal Interview NotesTransformer, LLM Architecture, Training/RLHF/DPO, Multimodal (CLIP/LLaVA/GPT-4o), Engineering PracticeOnline Reading
RL for LLM Alignment Quick ReferencePolicy Gradient → PPO, GRPO, RLHF Pipeline, DPO Derivation, Algorithm ComparisonOnline Reading
VLM Knowledge & Interview Guide (2025-2026)Visual Encoder, VLM Architecture, Alignment Training, Resolution Strategy, MoE, Latest ProgressOnline Reading
NLP & LLM Course NotesTokenization, N-gram, Embeddings, Neural LM, Transformer, GPT, BERT, RLHF (Fudan CS40008)Online Reading