How Large Language Models Supercharge Douyin’s User Experience

Background

With user growth plateauing and competition for attention intensifying, Douyin shifts its focus from acquiring users to optimizing the user experience. The platform introduces a unified "experience middle‑platform" that integrates signal identification, content understanding, problem attribution, and governance into a single capability chain.

Signal Recognition

Both explicit (feedback forms, reports) and implicit signals (posts, comments, search behavior) are collected. Multi‑modal models and anomaly clustering are used to recognize negative feedback at scale, converting raw user actions into structured signals for downstream processing.

Content Understanding

Large models generate concise feedback summaries, classify feedback into predefined categories, assign quality scores, and extract semantic viewpoints. Specialized pipelines (e.g., SFT‑fine‑tuned models, reward‑optimized models) improve summarization accuracy and enable automatic quality ranking of user reports.

Attribution and Diagnosis

Automated attribution combines rule‑based knowledge bases with LLM reasoning to pinpoint root causes across product, policy, and system layers. Experiments such as A/B tests are linked to feedback spikes, and LLMs generate diagnostic explanations that guide remediation.

Governance and Execution

Identified issues trigger automated or manual remediation actions. The system orchestrates cross‑functional teams (product, strategy, policy) to apply fixes, update rules, and close the loop with knowledge‑base updates. Continuous monitoring and feedback loops ensure iterative improvement.

Future Outlook

The roadmap aims to move from post‑event remediation to real‑time monitoring and pre‑emptive modeling, eventually integrating an AI‑driven agent framework that automates detection, understanding, attribution, and mitigation in a closed loop.