01
text · image · video
Multimodal content understanding
At Meta, I work on LLMs that understand content across language, images, and video—connecting model behavior to real product experiences.
Meta · Multimodal LLMs
Huiyu (Yvette) ChenMachine Learning Engineer at Meta
Hi, I’m Yvette. I build multimodal LLMs that understand content.
I’m a machine learning engineer at Meta, based in Singapore. My current work focuses on multimodal content understanding across language, images, and video—plus the training, evaluation, and production systems that make it useful.
research brain, product hands
Now · 现在
Building multimodal content-understanding LLMs at Meta—across language, images, and video—while staying interested in the whole path from how a model learns to how a person experiences it.
Full profile ↗︎What I do · 研究方向
Model work, systems work, and the bridge between them.
I’m happiest in the messy middle—where a promising idea has to become a reliable, measurable experience.
02
train · evaluate · ship
Large language model systems
I care about training, alignment, evaluation, and production systems that stay measurable, efficient, and understandable at real scale.
03
rank · retrieve · generate
Generative recommendation
I study how retrieval, ranking, semantic IDs, and language models converge—currently documented through a twenty-paper bilingual series.
A short timeline · 简历
The current chapter,
plus just enough backstory.
details belong on LinkedIn 〰
Now
Machine Learning Engineer · Meta
Building multimodal content-understanding LLM systems across text, images, and video.
Previously
Senior Machine Learning Engineer · Shopee
Built production AI assistants for e-commerce across multiple markets. That chapter taught me how to carry an LLM idea from training to real users—without turning this homepage into a quarterly report.
Before that
Computer Science · CASIA
Research training in NLP and machine learning, with a lasting habit of reading the appendix.
Off screen
Bouldering, tennis, hiking, swimming—and too many stairs.
Still learning, still moving, still evolving.
Speaking · 演讲
One stage, one production LLM story, twenty-seven slides.
A practical talk about taking chatbot assistants from architecture decisions and model training to alignment and production scale.
Open 27 slides ↗︎APAC Data Innovation Summit 2026
LLM-Powered Chatbot Assistants: Elevating the Customer Journey
I shared the system and model choices behind production-grade conversational AI—from retrieval and fine-tuning to alignment, evaluation, and deployment.
Topics · 专题入口
Follow a question, not just a feed.
长专题、前沿增刊和可以直接动手的实验，都从这里进入。 每个专题有自己的路线，不会被最新文章流冲走。
01
MODEL LEARNING · 双语学习路线
从第一个 token 学懂大模型
从张量和 Attention，到训练、模型输入输出、RAG 与前沿推理；真实模型与可编辑 Python 陪你动手。
02
MODEL LEARNING · 双语学习路线
LLM Infra：从显存到吞吐
五大训练/推理框架、GPU/TPU、kernel、分片与 profiling，把性能问题变成可验证的实验。
03
MODEL LEARNING · 双语学习路线
多模态内容理解
从图像 patch、CLIP 和 masked reconstruction，到 VLM、视频采样与时间定位。
04
MODEL LEARNING · 双语学习路线
HSTU：从行为到推荐
数据、逐张量推导、微型训练与规模化；连接多模态内容表示和序列推荐。
05
FOUNDATION SERIES · 基础主线
20 篇论文读懂生成式推荐
从 BPR 的打分器，到 OneReason 的推荐推理：沿着同一个问题，读懂排序、序列、语言任务、Semantic ID 与统一生成。
06
SEASON 2 · 2026 前沿
生成式推荐进入深水区
从可学习 ID、级联推理和连续 Token，到页面生成、广告部署与泛化审计：追踪 2026 年真正改变问题边界的七篇论文。
07
INTERACTIVE LAB · 动手实验
亲手走一遍生成式推荐
从 BPR 到 Semantic ID、beam search 与 Trie 约束：可在页面里改 Python，也可下载 Notebook 或直接进入 Colab。
Latest writing · 最新文章
Elsewhere · 生活支线
There is life outside the terminal, allegedly.
A spinning planet of places I’ve been, plus the less polished thoughts that escape onto Xiaohongshu.