Peking University
2026 — PresentMPhil in Computer Science
Advised by Prof. Daquan Zhou
About me
Research & engineeringI am Xu Huang, an MPhil student at Peking University, advised by Prof. Daquan Zhou. I am now interested in building useful AI that can act, evaluate its own work, and build complex systems through iterative, vision-in-the-loop reflection. I envision systems that can even build their own benchmarks, identify their own weaknesses, and create targeted environments and data to continually improve themselves.
In the pre-GPT era, I won several programming competition awards, including an ACM-ICPC Gold Medal, a CCPC Gold Medal, and an NOI Silver Medal.
(Fun fact: In the pre-agent era, I was the technical lead at an AI application startup that secured funding from MiraclePlus (奇绩创坛). Along the way, I gained hands-on experience with frontend and backend development, AWS infrastructure and deployment, a range of databases, and edge functions.)
Education
Beijing University of Posts and Telecommunications
2022 — 2026BSc Computer Science and Technology
GPA: 93.3 / 100 (Rank: 4 / 388)
Experience
Meituan — Beidou Program (美团北斗计划)
Feb 2026 — PresentResearch Intern
I am a major contributor to the development of discrete unified representations for the LongCat-Next and LongCat-Pro-V series.
I am also a core member of the team working on visual agent post-training for foundation models.
BAAI (Beijing Academy of Artificial Intelligence)
Jun 2025 — Feb 2026Research Intern
Participating in the development of the next-generation Emu unified multimodal large model.
Major contributor to the video generation component of the unified model.
Major contributor to the video generation component of the unified model.
Tencent
Feb 2024 — Jan 2025Research Intern
Led a text/image-to-3D-layout project, controllably constructing 3D scenes from a predefined asset library.
Tencent
Jul 2023 — Sep 2023Algorithm Engineering Intern
Utilized Score Distillation Sampling (SdS) and differentiable rendering to train diffusion models for SVG vector generation.
GitHub link
GitHub link
Project Experience
02 projectsVisual Agent Post-Training — Meituan
2026- Primarily responsible for visual agent post-training. Built task environments for PPT creation, HTML generation, interactive web development, and open-ended 3D scene construction using Blender and Three.js, and synthesized task data for post-training.
- Developed evaluation rubrics for functional usability and visual aesthetics to guide evaluation and post-training.
- Implemented automated, iterative optimization of agent harnesses and distilled effective harness strategies into the model.
Vision Encoder Training — Meituan
- Major contributor to the development of highly compressed discrete visual representations for unified understanding and generation, validated internally and adopted in LongCat-Pro-V (1.6T) and the LongCat-Next series.
- Systematically studied how scaling multi-level codebooks narrows the fidelity gap between discrete and continuous representations at high compression ratios, and how to integrate high-level semantic features with low-level pixel details.
Publications
05 worksEmu3.5: Native Multimodal Models are World Learners
Mainly contributed to the diffusion decoder, tokenizer, and Discrete Diffusion Adaption. Emu3.5 is one of the most advanced open-source native multimodal models developed by BAAI.
Read paper DC-SAE: Deep Compression Semantic Autoencoder for Faster Diffusion Convergence
Combines semantic features with a pixel-detail branch to achieve faithful reconstruction at 32× spatial compression and faster diffusion-model convergence. Code
Read paper SAE (Semantic AutoEncoder)
High-channel autoencoder built on DINOv3/SigLIP2 semantic representation encoders; end-to-end encoder/decoder training with geometry-aware regularization and a refined training recipe to balance semantic understanding, generation quality, and reconstruction fidelity.
SAE-T2I — Project Lead. A scaling-oriented extension of SAE for text-to-image generation, using SAE as the autoencoder backbone for high-quality image synthesis.
Imaginarium: Vision-guided High-Quality Scene Layout Design
- Used an image generation model to create scene images, segmented and analyzed the images, and retrieved relevant assets from a predefined library.
- Trained a DINOv2 model to estimate the 6D pose and scale of each object.
- Fine-tuned the image generation model using rejection sampling to improve similarity between generated images and the asset library.
Reviews: 3 / 3 / 3 / 1 (all positive).
QUITO-X: A New Perspective on Context Compression from the Information Bottleneck Theory
Used a small model as a proxy for a large model, using attention scores to identify key tokens during inference and compress redundant tokens to reduce the large model’s inference cost.
Read paper Honors & awards
- CCSP (Collegiate Computer Systems & Programming Contest) – 6th NationwideCCFOct 2023
- ACM-ICPC Gold MedalInternational Collegiate Programming ContestOct 2023
- CCPC Gold MedalChinese Collegiate Programming ContestNov 2023
- NOI Silver MedalNational Olympiad in InformaticsJul 2021