July 27th news: Today, the Douyin Group issued a notice stating that in order to alleviate parents' concerns about supervision and respond to the growth needs of minors, the platform has comprehensively upgraded the core engine of the minor mode - the age-appropriate recommendation algorithm. Once users activate the minor mode and fill in their birth date, this specialized age-appropriate recommendation algorithm will be activated.
This round of comprehensive upgrade of the minor mode recommendation algorithm focuses on introducing cutting-edge multimodal large language model (MLLM) technology, thus building an intelligent and refined content identification and distribution system. Multimodal large language models inherently possess cross-modal semantic understanding capabilities, capable of processing text, images, audio, and other types of information simultaneously, while also considering dimensions such as cognitive complexity and value orientation of the content, determining the appropriate age range for each video, and laying a technical foundation for subsequent age-appropriate recommendations.
The process works in several steps: First, the platform uses a high-quality multimodal large model for minors to screen a massive amount of information, selecting candidate content suitable for minors to watch; then it is handed over to human experts for strict review, selecting high-quality content with positive values, and including them in the exclusive content pool for minors; subsequently, the age-identification multimodal large model deeply understands and classifies these reviewed high-quality contents, outputting the appropriate age range information.
The official also mentioned that the platform currently adopts a detailed age classification standard for age-appropriate recommendations of high-quality content. This mechanism can adapt to the cognitive characteristics and psychological changes of minors at different stages, making the granularity of content supply more delicate; meanwhile, the algorithm will also recommend content based on the child's preferences. This "human screening + large model classification + age-appropriate recommendation + content preference matching" multi-layered guarantee, within the scope of the information the platform can collect, tries its best to make each recommendation more in line with the child's cognitive development needs; children can also give real-time feedback by long-pressing and clicking the "Not interested" button.
