name: Yuxiang Wang (Stein)
role: M.S. Student @ Fudan University
advisor: Prof. Baojian Zhou · Knowledge Works Lab
internship: Research Intern @ Alibaba (AI4S / Multimodal LLM)
location: Shanghai, ChinaResearch Interests
"The most interesting problems live at the intersection of paradigms — where autoregressive meets diffusion, where language meets vision, where training meets inference."
- LLMs & Agents — post-training, tool use, multi-turn reasoning (Thought → Action → Observation), RL for agents
- Diffusion Language Models — locality-aware architectures, training stability, inductive-bias / task-structure relationship
- AI4S — protein / small-molecule modeling, unified evaluation, Protein Agents
- Multimodal LLMs — unified understanding & generation, protein ↔ text bidirectional modeling
- Graph Neural Networks — dynamic-graph reasoning, fraud detection on spatio-temporal graphs
- 🧪 Diffusion Language Models — studying trainability of Masked Diffusion LMs vs AR across structured tasks
- 🤖 Agent RL — post-training agents with tool use and multi-turn reasoning
- 🧬 AI4S — protein language modeling and unified evaluation frameworks
Interview prep and study notes for LLM / Multimodal / RL. Feel free to ⭐ star and use!
| Resource | Description |
|---|---|
| 🔥 Agentic RL 全景指南 | PPO/DPO/GRPO/OPD 四大族谱 · 8 大 Part · 89 篇 arXiv · DeepSeek V4 / GLM-5.2 / Kimi K2.6 · ICLR/ICML 2026 — 前置基础见 RL for LLM Alignment |
| RL for LLM Alignment | Policy Gradient → PPO → GRPO, DQN/SAC, RLHF Pipeline, DPO Derivation — 进阶 Agentic RL 算法全景见 Agentic RL 全景指南 |
| 🔥 DLM Survey — 扩散语言模型前沿调研 | D3PM → LLaDA: 四大路线全景、Scaling Laws、投机解码、ELBO/RL 训练，50+ 核心论文 |
| Agent Harness 深度解析 | Agent 训练基础设施全景 — 四层架构、异步 Rollout、GRPO、VERL/ROLL/RAGEN 对比 |
| Transformer Decoder 全流程 | 训练(并行) vs 推理(自回归+KV Cache)、张量维度推导、RoPE/Attention/MLP 维度链 |
| Resource | Description |
|---|---|
| LLM 算法岗面试题库 (420+ 题) | Transformer / RLHF / RL / Agentic RL / VLM / Agent / RAG / Infra / 手撕代码 |
| LLM & Multimodal Interview Notes | Transformer, LLM Architecture, RLHF/DPO, Multimodal (CLIP/LLaVA/GPT-4o) |
| VLM Knowledge & Interview (2025-2026) | Visual Encoder, VLM Architecture, Alignment, Resolution, MoE |
Last updated: July 2026 · Design system by stein-wang0226.github.io