Panwang Pan profile photo

Panwang Pan | 潘攀望

Hi! I'm Panwang Pan, a senior researcher at ByteDance Seed. I work on multimodal agents for real-time perception, interaction, tool use, and feedback-driven improvement.

My through-line is RSI for real-world agents: perceive continuously, interact with people and tools, observe outcomes, and use feedback to refine behavior. I am especially interested in:

RSI loop: perceive, interact, use tools, and improve from feedback

I previously worked at Alibaba Cloud and interned at NVIDIA. I maintain open-source agent systems via PaulClawX and review for major ML/CV conferences. Always open to discussion and collaboration. Email: paulpanwang@gmail.com; WeChat ID: PaulClaw.

Email | Scholar | GitHub | PaulClawX | X | WeChat: PaulClaw | 小红书

📑 Selected Works [Google Scholar]
SeedRealtime: Real-Time Audio-Visual Agent for Natural Interaction

[Project] [ModelCard] [Blog]

An industrial real-time assistant that watches, listens, speaks, and responds in context, anchoring for agents that stay present in live environments instead of only answering turn-by-turn prompts.
Interactive & Proactive Agent

[Project] [Paper] [Code] [Dataset] [Demo]

proactive monitoring · task management · reactive-proactive interaction

IPI-Agent is training-free: interaction-control policy and temporal gating make existing offline MLLMs more stable, stateful, and timely in streaming assistance.

ICLR 2026

 SAM-Veteran: An MLLM-Based Human-like SAM Agent for Reasoning Segmentation

[Project] [Paper] [Poster]

An MLLM-based SAM agent that performs reasoning segmentation in a human-like workflow: it proposes boxes, refines masks with interaction points, and terminates adaptively through CoT-guided tool use.

NeurIPS 2025

JarvisArt demo
JarvisArt: Liberating Human Artistic Creativity via an Intelligent Photo Retouching Agent

[Project] [Demo] [ModelCard] [Dataset] [Code] JarvisArt GitHub stars

An MLLM-driven photo retouching agent that understands user intent, follows CoT-style artistic reasoning, and coordinates 200+ Lightroom tools through an Agent-to-Lightroom protocol to improve image quality with fine-grained global and local control.
🛠 Open RSI Lab [Home]

I build RSI systems where agents perceive continuously, interact with people, software, and devices, observe outcomes, and use feedback to improve future behavior.

Digital-world RSI workflow with browser, terminal, documents, phone, and feedback traces Digital-World RSI Agents use search, query, browser, computer-use, phone, CLI, and visual understanding tools, then convert digital outcomes into reusable feedback. Real-world RSI scene with smart glasses, autonomous driving, robotics, open environments, and feedback traces Real-World RSI Agents connect smart glasses, vehicles, robots, sensors, phones, and computers, then learn from feedback across open-world environments.
🌍 Agent Environments

World models provide the environment layer for capable agents: dynamic scenes, long-horizon memory, identity-consistent generation, and evaluation loops that connect simulated behavior to real-world action.

NeurIPS 2025 DynamicVerse

Scene dynamics for feedback-rich agent environments.

Project Code DynamicVerse GitHub stars
ECCV 2026 Ring Forcing

Long-horizon visual memory for agent rollouts.

CVPR 2026 Diff4Splat

Controllable dynamic scene generation from a single image.

Project Paper Code Diff4Splat GitHub stars CVPR 2026 MoVieS

Motion-aware dynamic view synthesis for fast agent-world rollouts.

Project Paper MoVieS GitHub stars CVPR 2026 ID-Crafter

Compositional multi-subject video generation for identity-consistent scenarios.

Project Paper Code NeurIPS 2025 PartCrafter structured scene generation overview PartCrafter

Compositional structure for scene generation and world-model environments.

GitHub PartCrafter GitHub stars
💼 Selected Experience
  • ByteDance PICO/Seed 08/2022 - Present, Senior Researcher, World Model & Multimodal Agents
  • Alibaba Cloud 07/2019 - 07/2022, Senior Researcher, Cloud Intelligent Computing and Embodied AI
  • NVIDIA 07/2018 - 10/2018, Intern, Computer Vision Research
💬 Professional Activities
  • Conference Reviewer: ICLR, NeurIPS, ICML, CVPR, ICCV, ECCV
  • Top Reviewer: NeurIPS 2025, ICLR 2025, ICML 2026
🌱 Hobbies Beyond Work
  • Football: Messi fan · Argentina supporter · Barcelona supporter · Paris Saint-Germain supporter · Inter Miami supporter
  • Also into: badminton, Go, and marathon running.