Conversational AI Large Model Researcher - Global E-commerce - Soaring Star Talent Program

Role Overview

This position sits within a team focused on creating a leading large-model dialogue system for e-commerce. The solution supports a very large user base across the Douyin commerce ecosystem and is designed to improve service efficiency and customer experience across multiple support scenarios.

Team Focus

The group develops intelligent dialogue capabilities for business use cases such as platform customer support, merchant support, creator or influencer support, and smart shopping assistance. Through ongoing technical iteration, the team has built an end-to-end conversational service solution that helps streamline operations and deliver better user experiences.

Research Goal

The main objective is to build an LLM-powered customer service chatbot for TikTok and Douyin e-commerce. The system should be able to manage the full inquiry journey, from understanding the request to negotiating a resolution and carrying out the final action.

Why This Work Matters

Because large language models are strong at dialogue and reasoning, they are well suited to customer service scenarios where responses must be both accurate and natural. The aim is to approach the quality of top human service representatives.

Key Research Directions

• Design a multi-agent architecture built on an LLM, combining planning, reply, and tool-use agents so the service flow can move from issue detection to solution negotiation, execution, and follow-up.
• Ensure the reply agent recommends actions that follow platform policies and service rules, while avoiding unsupported or inaccurate answers and maintaining smooth user communication.
• Have the planning agent identify what the user needs, determine the relevant service rules and constraints, and flag risk-related situations.
• Enable the tool agent to check whether actions are allowed, interpret tool outputs correctly, and manage dependencies between different operations.

Research Challenges

• Policy compliance: solutions must stay aligned with service rules, such as refund windows after parcel delivery and coupon usage limits per user per week.
• Adaptive interaction: instead of relying only on fixed rules, the system should respond dynamically to user behavior, current requests, and feedback so it can deliver more personalized service.
• Self-reflection: the model should be able to examine and improve its own decisions, especially when handling complicated or ambiguous tasks.
• Complex image understanding: the role includes dealing with varied and information-heavy images such as shipping records, bank payment screenshots, damaged-item photos, and seller qualification documents.

Company Overview

ByteDance was founded in 2012 with a mission to inspire creativity and enrich life. Its product ecosystem includes TikTok, Lemon8, CapCut, Pico, and China-market platforms such as Toutiao, Douyin, and Xigua.

Why Join ByteDance

The company emphasizes innovation, rapid iteration, curiosity, humility, and a strong focus on impact. Teams work in an environment shaped by an