Naver is positioning itself in the generative artificial intelligence (AI) competition by emphasizing its unique identity. The company aims to create an AI search service that seamlessly integrates user-generated content (UGC) from blogs and cafes, shopping, and local services within a single portal ecosystem.
Naver's goal is to combine search data with AI to better understand user inquiries and facilitate actual service usage, such as shopping and reservations.
On July 2, Naver held a Tech Deep Talk session in Gangnam, Seoul, where it unveiled its next-generation language model and core technologies for the AI search service, known as 'AI Tab.'
The newly introduced model is a lightweight version of the existing HyperCLOVA X (HCX), specifically developed for AI search. While general AI models focus on diverse knowledge and reasoning abilities, the Product Native LLM is designed to understand user questions and effectively utilize Naver's services, including search, blogs, places, and shopping.
Lee Ki-chang, Director of Naver Cloud, explained, "While the existing HyperCLOVA X was a general AI, the next-generation model combines general reasoning capabilities with service execution abilities. It can reliably perform AI agent functions without significantly increasing response times, even during lengthy conversations."
The new model also improves response speed and throughput. Unlike the previous model, which experienced significant delays with longer questions, the new design maintains stable performance during extended dialogues. The training data has expanded from general educational content to include specialized knowledge such as court rulings and academic papers.
From Memorizing Answers to 'Thinking AI'
The learning approach has changed as well. Previously, the model relied on rote learning of correct answers, but the next-generation model actively employs reinforcement learning to discover better response methods through trial and error.A notable example is 'clarity reinforcement learning.' Instead of generating plausible answers when uncertain, the model learns to ask users follow-up questions.
For instance, when asked, "Who is the lead actor in the educational drama?" it can provide an answer, but for a vague question like, "Who is the lead in that drama?" it responds with, "Which work are you referring to?" This behavior of asking additional questions when information is lacking is rewarded during training.
Naver has also strengthened its integration with its services. If a user requests, "Recommend a nice restaurant in Gangnam," the AI utilizes place search results. If the request specifies, "A place for two at 7 PM in Sinsa-dong," it checks reservation availability before generating a response. This approach aims to create an AI that not only provides answers but also selects necessary tools to synthesize results.
Naver introduced 'Harness Engineering' as a core technology for AI Tab. Instead of relying on a single large language model to perform all tasks, the service operates by combining optimized small models for specific tasks with various tools for search and reservations.
Han Seung-kyun, Naver's AI search leader, stated, "It is challenging to provide stable services with just one LLM. Designing which models and tools to use is the key to competitive AI services."
Naver claims this approach has reduced equipment costs by about three times compared to previous methods and improved response speed by more than double. The performance evaluation for service execution capabilities also surpassed the average of competing models.
Next Steps: From Text to Images
Naver plans to evolve AI Tab into a smart lens-based multimodal AI agent. Users will be able to show photos or videos, and the AI will understand the scene and context, enabling it to perform actions such as searching, shopping, and making reservations.For example, if a user sees a cafe in a video and requests, "Please reserve a similar cafe in my neighborhood for four people at 7 PM," the AI will understand both the image and text to complete the reservation.
Yoon Sang-du, a Naver leader, emphasized, "Our goal is to create AI that not only understands images but also allows users to see, comprehend, and execute actions. We aim to implement a multimodal AI that connects seamlessly with Naver services."
Meanwhile, Naver has stated that there are currently no plans to introduce advertising for AI Tab. The company explained that securing user trust is more critical at this stage than monetization. However, it left open the possibility of considering advertising in the future based on service maturity and business decisions.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.