Key skills
RAILLMLarge Language ModelsCommunicationCollaboration
About this role
TT-Search Algorithm & Applied AI is the algorithm team behind the search business built on TikTok (TikTok Search), with the goal of becoming the search engine of choice for users worldwide. Compared with recommendation systems, which passively infer user intent, search delivers content based on users' discovery motivation — making intent expression far more precise. This search data can also feed back into the recommendation engine, helping users obtain more relevant content.
We are at an inflection point where the search paradigm is shifting from "retrieval-and-ranking" toward "Agents proactively completing tasks." With large language models and Agents as our two driving wheels, the team builds the search-domain LLM foundation, the Agent execution framework (Harness), multi-agent collaboration, and self-improvement closed loops. We support the implementation of business scenarios such as multimodal AIGC creation, visual search, on-device intelligence, and long-horizon task Agents — spanning POI search, Wish search, automated evaluation, infrastructure, and more.
As a project intern, you will have the opportunity to engage in impactful short-term projects that provide you with a glimpse of professional real-world experience. You will gain practical skills through on-the-job learning in a fast-paced work environment and develop a deeper understanding of your career interests.
Applications will be reviewed on a rolling basis - we encourage you to apply early.
Job Responsibilities
- Search LLM Foundation R&D: Build and optimize the search-domain LLM foundation, integrating search knowledge to rapidly deliver business value; develop and refine the post-training pipeline for search LLMs (ultra-long-text / colloquial-text pre-training, image-text / video multimodal representation, e-commerce product multimodal representation learning, etc.); participate in LLM inference optimization (long-context optimization, model efficiency optimization).
- Agent Engineering & Multi-Agent Orchestration (Harness / Loop): Contribute to building the search Agent execution framework (Harness) — unified orchestration of tool calling, planning, memory, and environment interaction; multi-agent cluster scheduling and collaboration algorithms (task allocation, dynamic scheduling, inter-agent communication / alignment / conflict resolution); build the Agent Loop and the "training–inference–evaluation" closed-loop engineering.
- Long-term Memory, Self-Improvement & Data Closed Loop (Self-Evolve / RSI): Work on long-term memory mechanisms for LLMs, cross-context knowledge integration, and related directions; engage in cutting-edge exploration of self-evolve / self-improvement and recursive self-improvement (RSI) (automated hyperparameter tuning, training pipeline automation, AI-assisted algorithm design, model iteration closed loop); data synthesis and quality control (high-quality vertical-domain data synthesis, distribution alignment, synthetic data quality evaluation / filtering / refinement).
- Evaluation & Reward/Verifier Systems: Help build online Reward/Verifier systems and automated evaluation frameworks — annotation-free automated evaluation, evaluation of long-cycle complex tasks and cross-domain capabilities, and multi-agent collaboration evaluation standards.
- Business Implementation:
- Long-horizon task Agents (persistent intent)
- Multimodal AIGC creation: leveraging SOTA models for image/video generation to power the Feed's "ask-after-viewing / create-after-viewing" experiences and strengthen users' proactive mindset;
- Visual search & on-device intelligence: object detection, OCR, TinyLLM;
- Search content / creator ecosystem, etc.
We are at an inflection point where the search paradigm is shifting from "retrieval-and-ranking" toward "Agents proactively completing tasks." With large language models and Agents as our two driving wheels, the team builds the search-domain LLM foundation, the Agent execution framework (Harness), multi-agent collaboration, and self-improvement closed loops. We support the implementation of business scenarios such as multimodal AIGC creation, visual search, on-device intelligence, and long-horizon task Agents — spanning POI search, Wish search, automated evaluation, infrastructure, and more.
As a project intern, you will have the opportunity to engage in impactful short-term projects that provide you with a glimpse of professional real-world experience. You will gain practical skills through on-the-job learning in a fast-paced work environment and develop a deeper understanding of your career interests.
Applications will be reviewed on a rolling basis - we encourage you to apply early.
Job Responsibilities
- Search LLM Foundation R&D: Build and optimize the search-domain LLM foundation, integrating search knowledge to rapidly deliver business value; develop and refine the post-training pipeline for search LLMs (ultra-long-text / colloquial-text pre-training, image-text / video multimodal representation, e-commerce product multimodal representation learning, etc.); participate in LLM inference optimization (long-context optimization, model efficiency optimization).
- Agent Engineering & Multi-Agent Orchestration (Harness / Loop): Contribute to building the search Agent execution framework (Harness) — unified orchestration of tool calling, planning, memory, and environment interaction; multi-agent cluster scheduling and collaboration algorithms (task allocation, dynamic scheduling, inter-agent communication / alignment / conflict resolution); build the Agent Loop and the "training–inference–evaluation" closed-loop engineering.
- Long-term Memory, Self-Improvement & Data Closed Loop (Self-Evolve / RSI): Work on long-term memory mechanisms for LLMs, cross-context knowledge integration, and related directions; engage in cutting-edge exploration of self-evolve / self-improvement and recursive self-improvement (RSI) (automated hyperparameter tuning, training pipeline automation, AI-assisted algorithm design, model iteration closed loop); data synthesis and quality control (high-quality vertical-domain data synthesis, distribution alignment, synthetic data quality evaluation / filtering / refinement).
- Evaluation & Reward/Verifier Systems: Help build online Reward/Verifier systems and automated evaluation frameworks — annotation-free automated evaluation, evaluation of long-cycle complex tasks and cross-domain capabilities, and multi-agent collaboration evaluation standards.
- Business Implementation:
- Long-horizon task Agents (persistent intent)
- Multimodal AIGC creation: leveraging SOTA models for image/video generation to power the Feed's "ask-after-viewing / create-after-viewing" experiences and strengthen users' proactive mindset;
- Visual search & on-device intelligence: object detection, OCR, TinyLLM;
- Search content / creator ecosystem, etc.