PhD student Nanjing University

Tianyidan Xie.

谢天异丹 Sealical

Multimodal AI that can
generate, reason, and remember.

I'm a PhD student at Nanjing University, studying multimodal generation, agents, and spatial memory.

Previously at Skywork AI / Kunlun Tech, I contributed to UniPic, R1V, and Matrix-Game.

Email Google Scholar GitHub

Latest · Sep 2026LTE / ST-Mem — accepted to NeurIPS 2026.Explore the project

01 / Selected work

Research.

All publications on Scholar

LTE connects linguistic, spatial, and visual evidence into object trajectories across a home.

NeurIPS 20262026Accepted

Linguistic Trajectory Encoding for Efficient Long-Horizon Spatial Memory in Embodied Agents

Tianyidan Xie et al. First author

Object-centric memory that connects language, 3D space, and visual evidence for retrieval across hours of experience.

Paper Project Spatial memory

Updated PhysCodeBench overview: the 1,200-sample benchmark, three-agent SMRF framework, and evaluation of code, visuals, and physics.

Preprint · Under review2026

PhysCodeBench: Benchmarking Physics-Aware Symbolic Simulation of 3D Scenes via Self-Corrective Multi-Agent Refinement

Tianyidan Xie et al. First & corresponding author

A physics simulation benchmark and a self-correcting, three-agent framework for executable, physically grounded code.

Paper Project Agents for simulation

PhysLayer animates falling stones, a bouncing ball, a moving car, and a falling vase with depth-aware physics.

ICME 20262026

PhysLayer: Language-Guided Layered Animation with Depth-Aware Physics

Tianyidan Xie et al. First author

Bringing still images to life through language-guided, depth-aware physical simulation.

Paper Project Physics-aware generation

CineAGI generates movie scenes with consistent characters and coordinated audiovisual narratives.

ICME 20262026

CineAGI: Character-Consistent Movie Creation through LLM-Orchestrated Multi-Modal Generation and Cross-Scene Integration

Tianyidan Xie et al. First author

Multi-agent orchestration for coherent movie creation, with consistent characters and synchronized audio and video.

Paper Project Multi-agent creation

Anywhere places foreground objects into diverse, semantically consistent generated backgrounds.

AAAI 20252025

Anywhere: A Multi-Agent Framework for User-Guided, Reliable, and Diverse Foreground-Conditioned Image Generation

Tianyidan Xie et al. First author

VLM, LLM, and image-generation agents collaborate to create reliable and diverse backgrounds around a given foreground.

Paper Project Code Foreground-conditioned generation

EgoMonth connects daily first-person recordings across a month to evaluate long-term spatiotemporal memory.

Preprint · Under review2026

EgoMonth: A Month-Level Egocentric Video Benchmark for Long-Term Spatiotemporal Memory

Weitao Chen, Jiaxin Hu, Tianyidan Xie, et al. Co-author

Evaluating whether multimodal models can maintain spatiotemporal memory across weeks of everyday experience.

Paper Project Long-term video understanding

02 / Research in industry

Ideas into models.

Core contributions at Skywork AI
Kunlun Tech · 2025–2026

Unified generation01

Skywork UniPic

I contributed to UniPic 1.0 and 2.0, unifying image understanding, generation, and editing through autoregressive modeling and reinforcement learning.

One model. Multiple visual capabilities.

1.0 paper 2.0 paper Code

Multimodal reasoning02

Skywork R1V

I contributed to the R1V series, including R1V2's hybrid reinforcement learning and R1V4-Lite's interleaved visual reasoning, image operations, and multimodal search.

Think with images. Act with tools.

R1V R1V2 R1V4 Repository

Interactive world models03

Matrix-Game 3.0

I contributed to real-time world generation with camera-aware memory, supporting spatial and temporal consistency across long interactive sequences.

5B parameters · 720p · up to 40 FPS

Paper Project Code

Additional technical reports 01

Skywork-VL Reward: An Effective Reward Model for Multimodal Understanding and Reasoning2025

03 / Background

Along the way.

Past experience

Feb 2025 — Apr 2026

Skywork AI / Kunlun Tech

Large Model Researcher

Core contributor · UniPic, R1V & Matrix-Game
Dec 2023 — Feb 2025

Nanjing University

Research Assistant
Mar 2022 — Jun 2023

Xiaohongshu

NLP Algorithm Engineer
Jul 2021 — Jan 2022

Tencent Music Entertainment

NLP Algorithm Engineer · Intern

Education

Master's · Big Data Management · 2019–2022
Southwestern University of Finance and Economics

Bachelor's · Information Security · 2015–2019
Southwest University of Science and Technology

Keep in touch

Let's talk research.

Multimodal intelligence, visual worlds, and what comes next.

sealical@outlook.com

Tianyidan Xie.

Research.

Linguistic Trajectory Encoding for Efficient Long-Horizon Spatial Memory in Embodied Agents

PhysCodeBench: Benchmarking Physics-Aware Symbolic Simulation of 3D Scenes via Self-Corrective Multi-Agent Refinement

PhysLayer: Language-Guided Layered Animation with Depth-Aware Physics

CineAGI: Character-Consistent Movie Creation through LLM-Orchestrated Multi-Modal Generation and Cross-Scene Integration

Anywhere: A Multi-Agent Framework for User-Guided, Reliable, and Diverse Foreground-Conditioned Image Generation

EgoMonth: A Month-Level Egocentric Video Benchmark for Long-Term Spatiotemporal Memory

Ideas into models.

Skywork UniPic

Skywork R1V

Matrix-Game 3.0

Along the way.

Past experience

Skywork AI / Kunlun Tech

Nanjing University

Xiaohongshu

Tencent Music Entertainment

Let's talk research.