Job Description

Team Introduction:
Through algorithm optimization and collaboration with business teams, the team conducts comprehensive quality and ecosystem governance for ByteDance's e-commerce products. This involves combating risks, violations, and low-quality issues, as well as constructing and optimizing a healthy e-commerce ecosystem. The team aims to maximize platform governance effectiveness while improving operational efficiency and reducing costs. Additionally, the team is dedicated to advancing cutting-edge AI technologies to drive business transformation and development through technical innovation, covering diverse fields including but not limited to NLP, CV, multimodal models, large models, graph algorithms, and sequence algorithms.

Project Objectives and Rationale:
Intelligent content moderation in e-commerce is a highly complex field. As moderation technologies continue to evolve, various domains are encountering new risks and adversarial content, which pose fresh challenges for the application of foundation models. For example, existing open-source foundation models underperform in E-commerce moderation tasks involving PBR changes, long text, long sequences, multilingual content, few-shot scenarios, and AIGC-generated adversarial content. Consequently, there is an urgent need to develop foundation models specifically tailored for intelligent e-commerce moderation to improve their effectiveness and adaptability in e-commerce governance. In particular, we must explore high-quality data auto-generation, efficient MOE Embedding, Auto-prompt generation, high-quality COT output, and foundation model knowledge distillation. The model should also achieve high-accuracy autonomous decision-making and interpretable COT generation, significantly reducing misjudgments. For dynamic PBR changes, it should automatically retrieve similar moderation cases via RAG modules, decompose complex PBRs into simple atomic tasks, split rejection and exemption tasks, and auto-invoke corresponding tools, establishing an industry-leading intelligent review system that ""knows to reject and why."" Ultimately, the large language model-based intelligent moderation system should approach or exceed human moderators' accuracy and evolve toward fully automated review.

Project Content:
Research on e-commerce intelligent moderation multimodal large language models includes, but is not limited to:
Modality fusion: Enhance fine-grained understanding of text, audio, image, video, and live-streaming data to enable high-accuracy autonomous decision-making and interpretable COT generation.
Few-shot capabilities: Address e-commerce multilingual, long-sequence, and few-shot challenges, strengthen Few-Shot/Zero-Shot capabilities, and enable complex instruction and auto-prompt generation for dynamic business rules.
Adversarial defense: Study AIGC image/video discrimination to enhance the review model's ability to defend against vague and abstract generated content.
Agent capabilities: Enable RAG module invocation, tool usage, and Auto-planning; improve the model's dynamic reasoning and reflection abilities.

Involved Research Directions:
Large language models, multimodal large language models, Few-shot learning, AIGC decision-making, AIGC data generation, reinforcement learning, Agent