Peiyu Wang (王沛雨) Research Engineer & Tech Lead @ Skywork AI Large Language Models · Multimodal Systems · Agent Architectures peiyuwang.cs@gmail.com

I've been involved in the full pipeline of foundation model development — pre-training, supervised fine-tuning, RL alignment, evaluation, and production deployment. My work spans several directions:

• Video Generation: Co-developed SkyReels V4, a multimodal video model with 1080p/32FPS output and audio-video synchronization, trained with full-modal reinforcement learning.
• World Models: Built Matrix-Game 3.0, a memory-augmented interactive world model supporting 720p/40FPS real-time streaming with long-horizon consistency.
• Agent Systems: Developed SkyClaw-v1.0 and Super Agents, agent models with million-token context optimized for tool use, multi-turn execution, and code generation.
• Multimodal Reasoning: Contributed to the R1V Series (38B VLM) and UniPic Series (1.5B unified generation/understanding model).

SkyReels V4 — Multimodal video generation with full-modal RL. #1 on Artificial Analysis for text-to-video with audio.

Matrix-Game 3.0 — Memory-augmented interactive world model for real-time streaming video. Paper

SkyClaw-v1.0 — Agent model with million-token context for tool use and code generation.

R1V Series — 38B VLM with multimodal chain-of-thought reasoning. Paper

UniPic Series — 1.5B unified model for image understanding, generation, and editing. Paper

VL Reward Model — Multimodal reward model for RL alignment of MLLMs.

Super Agents — End-to-end agent system for autonomous task execution.

RED Recommendation — Large-scale ranking, retrieval, and multi-objective optimization.

• ML Summit 2026 Talk — Interactive World Models
• ML Summit 2025 Talk — Agent System Design
• Skywork Office AI Super Agent — Official Launch

Skywork Office AI Super Agent

SkyReels V4

• 量子位 — 杀进全球榜TOP2！国产视频模型黑马刚刚出现了
• 量子位 — 刚刚，全球视频模型新王诞生了！
• 机器之心 — 又一国产全模态视频大模型杀入Artificial Analysis榜单Top 2
• 新智元 — 刚刚，国产视频模型登顶全球第一

Matrix-Game 3.0

• Paper — Real-Time and Streaming Interactive World Model
• 量子位 — 国产玩家亮剑世界模型！天工AI不藏了

SkyClaw-v1.0

• 量子位 — 国产Agent模型闯入全球第一梯队
• 凤凰网 — 天工AI发布Agent模型SkyClaw-v1.0
• 网易 — 后OpenClaw时代，国产Agent凭「高配低价」叫板Opus 4.6

• Large-scale language models: pre-training, RL alignment, and evaluation
• Multimodal foundation models: reasoning, generation, and unified architectures
• World models: interactive simulation and robotic embodiment
• Agent systems: autonomous decision-making and tool orchestration

_{Built by SkyClaw v1.0 — the agent that writes resumes so you don't have to.}