Machine Learning Engineer Graduate (E-Commerce Content Recommendation - Generative & Large Recommendation Model) - 2027 Start (PhD)

Team Introduction 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 our team. As a graduate, you will get opportunities to pursue bold ideas, tackle complex challenges, and unlock limitless growth.

Successful candidates must be able to commit to an onboarding date by the end of the year. Please state your availability and graduation date clearly in your resume.

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.

- 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. - 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. - 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. - 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. - 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