01
Agent Harness
Harnesses tuned to one task rather than all of them, mixing models instead of betting on a single one: cheap work routed to cheap models, budget spent only where it changes the answer. The likes of oh-my-openagent and oh-my-pi.
Tokyo, JapanLocal--:--:--
杨思祺
SiqiYang
Ph.D. Candidate in Computer Science at the Camera Intelligence Lab, Peking University.
I work on spike cameras.video generation.world models.light & pixels.
01 — Research
Light, sampled differently.
Everything I work on is one question about light, asked three ways: how to recover a scene, how to capture one, and how to generate one that was never there.
Recovery is the inverse problem: intrinsic decomposition and inverse rendering prise geometry, material and light apart, and neural radiance fields do it in three dimensions.
Capture stops treating the sensor as given: a spike camera fires the instant a pixel has seen enough light, tens of thousands of times a second, which is what brings HDR and high-frame-rate imaging to scenes too fast and too contrasty for a shutter to hold.
Generation runs the arrow backwards: video generation that follows a camera and a soundtrack, relighting under a sun the scene never saw, world models consistent for everyone inside them.
02 — Interests
Where the spare cycles go.
What the evenings are for, and what keeps me honest about how much of this technology actually works in someone's hands.
02
Game Generation
Handing the last two years of LLM and coding-agent progress to players and makers rather than researchers, so anyone can build the game already in their head — at a cost they can predict and a quality they can count on.
03
Serverless Infra
Serverless primitives on Cloudflare, used to keep services small, cheap, and forgettable. I would rather pay per request than spend the next decade keeping a machine alive.
03 — Path
Where I have been.
- 2025 — Nowexperience
Alaya Lab, Shanda AI Research Tokyo
Research Intern · Japan
Multiplayer world models, panoramic video generation, generative game platform.
- 2022 — 2027education
Camera Intelligence Lab, Peking University
Ph.D. in Computer Science · Beijing, China
Advised by Prof. Boxin Shi and Prof. Zhaofei Yu.
- 2019experience
Hyperplane Lab, Peking University
Research Intern · Beijing, China
Convolutional neural network quantisation and pruning.
- 2018 — 2022education
Turing Class, Peking University
B.S. in Computer Science · Beijing, China
Honours track in the School of EECS.
- 2018experience
Youtu X-Lab, Tencent
Research Intern · Shenzhen, China
Object detection and classification on the Tencent Weishi dataset.
04 — Publications
Selected publications.
17 papers across CVPR, ICCV, ECCV, NeurIPS, SIGGRAPH Asia and TPAMI.
- 2026PAWA: Panoramic Action-conditioned World Modeling for Multiple Agents
- 2026SPW-Nav: A Streaming Panoramic World Model for Language-Guided Navigation
- 2026Programmable World Model
- 2026MASS: Multiplayer World Models with Authoritative Shared State
- 2026Generative World Renderer at the Speed of Play
- 2026High-Speed Full-Color HDR Imaging via Unwrapping Modulo-Encoded Spike Streams
- 2026InstructAV2AV: Instruction-Guided Audio-Video Joint Editing
- 2026MUGEN: Interactive Panoramic World Exploration via Camera Control
- 2026HFR and HDR Video from Multi-Attenuated Spikes Using a Rapidly Rotating SpokeND Filter
- 2026Audio-sync Video Instance Editing with Granularity-Aware Mask Refiner
- 2025PanoWan: Lifting Diffusion Video Generation Models to 360° with Latitude/Longitude-aware Mechanisms
- 2025SpikeDiff: Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-Millisecond Spike Streams
- 2025EventUPS: Uncalibrated Photometric Stereo Using an Event Camera
- 2024Real-Data-Driven 2000 FPS Color Video from Mosaicked Chromatic Spikes
- 2024Focal Stack and Light Field Imaging
- 2023Complementary Intrinsics From Neural Radiance Fields and CNNs for Outdoor Scene Relighting
- 2023MILO: Multi-bounce Inverse Rendering for Indoor Scene with Light-emitting Objects
05 — Recognition
Medals, mostly from problems.
- 2025PKU–Huawei Turing Graduate Program Scholarship
- 2025Huawei Hackathon, Software ChallengeGold Medal
- 2019Peking University Academic Excellence Award
- 2018China Collegiate Programming Contest, GuilinGold Medal
- 2017National Olympiad in Informatics (NOI)Silver Medal
- 2016–2017National Olympiad in Informatics in Provinces (NOIP)First Prize ×2
- 2016–2017APIO, CTSC, and Winter CampBronze Medals
Academic Service
Conference Reviewer
- CVPR
- ICCV
- ECCV
- NeurIPS
- SIGGRAPH
- ACCV
- PRCV
- ICIG
Journal Reviewer
- TPAMI
Toolkit
06 — Toolkit
Languages
Python · C++ · CUDA · Rust · TypeScript · Lisp · Nix
Deep Learning
PyTorch · JAX · TensorFlow · Diffusers · Triton
Graphics & Imaging
NeRF · 3DGS · Differentiable Rendering · Spike Sensor Pipeline
Infrastructure
Docker · Ray · Cloudflare · GCP · Git
07 — Open Source
Built for myself. Patched for everyone.
I build the tools I need, and send fixes upstream when the tools I use have rough edges — small, practical changes that keep shared software useful.
Own project
cloudflare-zotero-mcp
A remote Zotero MCP server on Cloudflare Workers, with full metadata and WebDAV PDF read/write, OAuth, and hybrid semantic search.
- TypeScript
- Cloudflare Workers
- MCP
- Zotero
- WebDAV
Upstream contribution
junhoyeo/tokscale
Added Jcode support, then fixed cached-token accounting and macOS configuration paths.
- Rust
- CLI
- Coding Agents
Own project
cloudflare-workers-mikan
A serverless mirror of Mikan on Cloudflare Workers, including account flows, custom domains and a live deployment.
- TypeScript
- Cloudflare Workers
- Hono
- Serverless
Upstream commit · 2019
WakeupSchedule_Kotlin
Fixed PKU course imports by skipping unselected rows, correcting week parsing and restoring the elective-system URL.
- Kotlin
- Android
- PKU
Learning project
PKU-Racket
Functional-programming exercises spanning SICP, continuations, interpreters, amb evaluation and register machines.
- Racket
- SICP
- Functional Programming
Upstream contribution
numtide/treefmt-nix
Added isort as a supported formatter, including its Nix module and checks.
- Nix
- isort
- Formatting
Learning project
SoftwareFoundations
Coq proofs and exercises from Software Foundations, covering induction, tactics, logic, proof objects and IMP.
- Coq
- Formal Methods
- Proofs
Upstream contribution
kaitranntt/ccs
Corrected Codex and Claude quota parsing across additional rate limits and utilization payloads.
- TypeScript
- Codex
- Claude
Learning project
PyTorch-FBS
A PyTorch implementation of Feature Boosting and Suppression for input-dependent dynamic channel pruning.
- Python
- PyTorch
- Dynamic Pruning
Upstream contribution
yuhp/opencode-models-discovery
Fixed models.dev matching for models served by custom providers.
- TypeScript
- OpenCode
- models.dev
08 — Contact
