About the role
Responsibilities
Global E-Commerce (TikTok Shop) is one of TikTok's fastest-growing businesses and a core driver of the company's revenue growth. Our Global E-Commerce Content Recommendation team owns the end-to-end recommendation stack for e-commerce video and image-text content on TikTok worldwide - retrieval, ranking, and multi-queue blending; supply ecosystem and cold start; and the browsing-to-purchase experience for hundreds of millions of users.
We believe recommendation is being rewritten in the compute era. ID-based collaborative filtering and supervised learning built today's systems and still run most of the industry - but their returns are diminishing, and we are betting the next order of magnitude on rebuilding the stack on LLM foundations. Our ambition is to build the most advanced recommendation system in the world, and the next generation after that.
We are looking for talented individuals to join us for an internship. PhD internships at Our Company provide students with the opportunity to actively contribute to our products and research, as well as to the organization's future plans and emerging technologies.
Our dynamic internship experience blends hands-on learning, enriching community-building and professional development events, and collaboration with industry experts.
Applications will be reviewed on a rolling basis, so we encourage you to apply early. Please clearly state your availability in your resume (Start date, End date).
Responsibilities
You will help build - and rewrite - an industrial recommendation system serving a billion-scale user base across short-video, livestream, and product scenarios, covering retrieval, pre-ranking, ranking, and blending end to end. Every iteration ships to production and directly moves user experience and GMV.
1. Scale recommendation models like LLMs. Push ranking models from hundreds of millions to billions of parameters and chart the scaling laws of recommendation: behavior-corpus pretraining; multi-scenario, multi-task, multi-stage joint training; ultra-long behavior-sequence modeling (10K+ events) with KV caching, sequence compression, user/generation (U-G) disaggregated serving, speculative decoding, and dynamic batching - raising MFU while holding a strict millisecond latency budget.
2. Build one-stage generative retrieval. Reframe retrieval as generation: tokenize the item space into semantic IDs (RQ-VAE / SID) and train autoregressive models, grounded in MLLM semantics, to generate what a user wants next - collapsing the traditional "multi-channel retrieval + ranking" funnel into a single generative stage. The open problems span the full stack: item tokenizers that balance semantic content against collaborative signal, and SIDs that stay stable while millions of new items arrive daily; post-training the generator directly on live user feedback (preference optimization, GRPO-style RL); and decoding under a millisecond budget - beam search, decoding constrained to the valid item space, and test-time scaling that trades inference compute for better recommendations. The prize is a system freed from its path dependence on ID memorization, where cold-start generalization comes from semantics rather than impression history.
3. Inject world knowledge. Use large models' real-world knowledge to mine latent user interests and semantic representations beyond what pure ID co-occurrence can express; use reasoning models to run explicit chain-of-thought inference over long-horizon user intent, making the system materially better at discovery and novelty.
4. Push training and inference to the hardware limit. Custom CUDA / Triton fused kernels, memory and computation-graph optimization, distributed training and inference acceleration, mixed precision and low-bit quantization - engineered for what makes recommendation hard: sparse embeddings, variable-length sequences, and many task heads.
5. Rewrite R&D with agents. We are embedding coding agents deep into the algorithm-development loop: automated feature mining and pipeline generation, experiment configuration and training orchestration, automated evaluation and online-diagnosis attribution, bad-case mining and patrol. You will be both a user and a builder of this system.
6. Do original work on open problems. Long-term value modeling, repurchase and retention, transaction attribution, fatigue modeling, new-user recommendation, incremental value modeling, interest exploration, LLM4Rec - problems where industry has no standard answers. We expect, and support, original research: internal papers, patents, and publication at top external venues.
Qualifications
Minimum Qualifications
- Currently pursuing a PhD in Computer Science, Electrical Engineering, Mathematics, Statistics or a related discipline.
- Solid ML and engineering fundamentals: you understand the math behind the models, and you write clean, efficient, reproducible code with a strong command of algorithms and data structures.
- Deep research or engineering practice in at least one of: LLMs / foundation models, NLP, CV, RL, or recommendation / search / ads - and you can articulate why you made the choices you made, and where they fell short.
- Genuine enthusiasm for LLM / LRM techniques: you want frontier methods live in production, not parked at offline metrics.
- Strong problem definition and decomposition: faced with an ambiguous problem that has no standard answer, you find your own foothold.
Preferred Qualifications
- Publications at KDD, SIGIR, RecSys, WWW, ACL, NeurIPS, ICML, ICLR, or comparable venues - or high-quality open-source work.
- CUDA / Triton kernel development, source-level deep-learning-framework optimization, large-scale distributed training, or high-performance inference deployment.
- Hands-on experience with LLM post-training (SFT / RLHF / DPO / GRPO), agent-system construction, or inference acceleration.
- Led or deeply contributed to a key project in search, ads, recommendation, or large models, with a complete problem-to-online-impact loop.
- Awards in ACM-ICPC, NOI, Kaggle, or comparable competitions.
- Heavy user of AI coding and agentic workflows for building systems and optimizing models.