Naver unveiled a new AI model on the 2nd that incorporates real-life information from mom cafe posts and restaurant reviews to improve answer accuracy. The model was introduced during a session titled 'From Search to Execution: Next-Generation AI Technology Creating Naver AI Search' held at Naver D2SF Gangnam. The lightweight model, based on HyperCLOVA X, applies 'harness engineering' to reflect life-oriented information found within Naver's ecosystem, allowing the AI to provide answers grounded in actual user experiences rather than generic data.
Naver Applies Harness Engineering to Integrate User-Generated Content
The AI model deployed in Naver's 'AI Tab' search service uses harness engineering, a structure that controls and directs AI agent behavior like horse tack guides a horse. When users search for math academy tuition, the system provides answers based on recent discussions by parents in mom cafes. When users inquire about restaurant parking availability, the AI reflects not only official business information but also actual customer reviews from people who have parked there.
Han Seung-kyun, leader of Naver AI Search Service, compared harness engineering to AI's 'work sense,' stating that "to create AI agents that work well in services, harness engineering that simultaneously designs cost efficiency and stability is essential, not just large language models (LLM)." Han added that "the experience of connecting search infrastructure and know-how accumulated over the past 27 years, vast content such as blogs and cafes, and various service assets such as shopping and places with AI technology to lead from search to execution is Naver's unique competitiveness that no one can easily follow."
Naver Reduces Operating Costs by 3x Using Small Language Models
Naver built a division-of-labor Small Language Model (SLM) structure to increase the efficiency of the AI Tab. Instead of one giant LLM handling all tasks, Naver combines role-specific SLMs to reduce operating costs while simultaneously improving response speed and quality. Using this method, Naver reduced equipment operating costs by up to 3 times and improved response speed by more than 2 times.
Naver trained the AI using Clarify RL (Reinforcement Learning) technology, which ensures the AI does not arbitrarily interpret ambiguous requests but confirms user intent through additional questions. As a result, hallucination decreased by up to 30 percentage points compared to HyperCLOVA X.
Lee Ki-chang, director of Naver Cloud HyperScale AI Model, emphasized that "the model applied to the AI Tab is characterized by optimizing the entire process from training data construction to model design and reinforcement learning to fit Naver services," adding that "the goal is to create a model that operates most accurately and efficiently in actual service processes such as search, purchase, and reservation by Naver users."
Naver Integrates SmartLens Multimodal Technology into AI Search
Naver disclosed its multimodal technology advancement strategy utilizing SmartLens placed at the front of the AI Tab search bar. Multimodal technology enables AI to understand and utilize various forms of information including images and videos, not just text.
Users can upload images instead of text in the search bar and issue commands such as 'find products similar to this' or 'book a place with this atmosphere.' While image and video search was possible with existing AI, LLM models first recognized images and then conducted text-based searches. In contrast, Naver's multimodal agent understands the entire process from image input to execution as a single request.
Yoon Sang-doo, leader of Naver Future AI Center, stated that "the visual search technology Naver has accumulated through SmartLens for nearly 10 years is core technology for creating eyes through which AI agents view the world," adding that "in the future, Naver's AI agent service will evolve in a direction that understands user intent not only through text but also through images and connects to actual actions."
FAQ
What did Naver unveil on the 2nd?
Naver unveiled a new AI model on the 2nd at Naver D2SF Gangnam that incorporates life-oriented information from mom cafe posts and restaurant reviews. The model is based on HyperCLOVA X and uses harness engineering to reflect real user experiences in its answers.
How much did Naver reduce operating costs with the new AI model?
Naver reduced equipment operating costs by up to 3 times using a Small Language Model (SLM) structure. The company also improved response speed by more than 2 times and reduced hallucination by up to 30 percentage points compared to HyperCLOVA X.
What is SmartLens multimodal technology in Naver's AI search?
SmartLens is Naver's multimodal technology that allows users to upload images in the search bar and receive answers to commands like 'find similar products' or 'book places with this atmosphere.' The technology understands the entire process from image input to execution as a single request, leveraging visual search technology Naver has accumulated for nearly 10 years.
