The posting
About the Team Live commerce may be the hardest recommendation problem at TikTok. The "item" is a live room hosted by a real person: the product on display, the price, the crowd and even the auction clock change by the second. Purchase signals are sparse, intent shifts in real time, and a single session can take a user from a short video to a live room to checkout in under a minute.
It is also a business growing at remarkable speed. Content-driven live commerce has already reshaped retail in another major market, turning a content platform into one of its top three e-commerce players within a few years. TikTok Shop is now scaling that model rapidly across the US and global markets, with live commerce at the heart of its growth.
The Global E-commerce Recommendation Live Algorithm team owns the end-to-end recommendation stack for TikTok Shop LIVE, from retrieval and pre-ranking to ranking, mixed ranking and long-term value modeling. One unified stack serves seven core surfaces, spanning public discovery and private follower relationships: For You feed live previews, creator-avatar LIVE entries, the swipe-up LIVE feed, LIVE discovery, Shop, Following and Inbox. We are rebuilding this stack around foundation models: large recommendation models that follow scaling laws, generative retrieval on multimodal semantic IDs, multimodal LLMs that bridge content and commerce, reward models and RL post-training that align recommenders with long-term value, and AI agents that tune, debug and evolve the system alongside us. What we ship directly drives GMV, buyer growth and creator success for TikTok Shop in the US and across global markets.
We are looking for experienced engineers and researchers who want to define the next generation of recommendation for live commerce, and see their ideas go live at TikTok scale.
Responsibilities: - Scale Large Recommendation Models (LRM): Push the scaling frontier of our unified ranking foundation model (model capacity, sparse MoE, ultra-long user sequences, pre-training and teacher-student distillation), and turn offline scaling gains into online wins under strict latency and compute budgets, with dynamic compute allocation and on-device real-time inference. - Generative recommendation on multimodal Semantic IDs: Design semantic IDs that fuse visual, speech, text, collaborative and product signals, including what is being showcased, pinned or auctioned in a live room right now, and power generative retrieval with near-real-time SID serving. - LLM4Rec and MLLM cross-domain modeling: Use multimodal LLMs to place live rooms, short videos and products in one semantic space; inject world knowledge to transfer user interest across content, search and shopping, and crack cold start for new creators, new products and new markets. - Reward models and post-training for recommenders: Build reward models for conversion, long-term value and user experience; align generative recommenders with SFT, DPO / GRPO-style RL post-training, generator-evaluator decoding and listwise objectives. - Agent x Recommendation: Build LLM agents that auto-tune multi-objective value weights and traffic mechanisms, design and read experiments, and diagnose online cases, moving from hand-tuned heuristics to agentic optimization and multi-step RL. - Long-term value and ecosystem: Model LTV, uplift and multi-touch attribution; design value and traffic mechanisms for live auctions, brands and emerging creators; balance short-term GMV with user experience and a healthy, growing creator ecosystem. - Own it end to end: From problem framing and data pipelines to large-scale training, online serving, A/B testing and launch, partnering closely with engineering, product and data science.
Minimum Qualifications - Bachelor's degree or above in Computer Science, Electrical Engineering, Mathematics, Statistics or a related field. - 3+ years of industry experience in recommendation, search, advertising or other large-scale applied machine learning, with a track record of models that shipped and moved online metrics. - Strong machine learning fundamentals and hands-on depth in at least one of: retrieval and ranking, sequential user modeling, multi-task learning, generative recommendation, or LLM / multimodal modeling. - Proficiency in Python and/or C++, and solid experience with PyTorch or TensorFlow, including distributed training and model optimization. - Ability to turn ambiguous business problems into modeling solutions, and to own iteration from offline experiments to online A/B results.
Preferred Qualifications - Experience in e-commerce or live-streaming recommendation, ads ranking, or other real-time, conversion-driven systems - Hands-on experience with large recommendation models or generative recommendation, including scaling laws, MoE, long-sequence modeling, pre-training or distillation. - Experience in LLM / MLLM post-training (SFT, reward modeling, RLHF, DPO, GRPO or other RL-based alignment), ideally applied to recommendation or ranking. - Experience building LLM agents (tool use, planning, automated experimentation or parameter tuning), or applying reinforcement learning to multi-step decision-making such as traffic allocation. - Experience with cross-domain transfer, LTV / long-term value modeling, uplift modeling and causal inference, or multi-objective optimization. - Experience with large-scale training and inference efficiency (mixed precision, model parallelism, GPU serving optimization). - Publications at top venues such as NeurIPS, ICML, ICLR, KDD, SIGIR, WWW, RecSys, ACL or CVPR, or strong results in major ML competitions.
