Naver has unveiled a product-native large language model (LLM) optimized for its conversational AI search service, AI Tab, along with a new harness engineering framework designed to improve reliability in real-world applications. The company plans to roll out these technologies as part of a major AI Tab upgrade in the second half of this year.
At the AI Search Tech Deep Talk Study event held in Seoul's Gangnam district on Wednesday, Naver introduced its next-generation AI model and the operational technologies powering AI Tab.
The new model is a lightweight version of HyperCLOVA X, Naver's proprietary LLM, tailored specifically for AI search. Unlike a general-purpose model, the product-native LLM was developed using Naver's own data, service scenarios, and user feedback throughout the design process.
Lee Ki-chang, Director of Hyperscale AI Models at Naver Cloud, said the model was built to handle long multi-turn conversations, select the most appropriate tools based on context, and complete users' tasks from start to finish.
The model adopts a customized Mixture of Experts (MoE) architecture, doubling response speed compared with its predecessor while minimizing latency for responses of up to 16,000 tokens.
Naver also incorporated clarity reinforcement learning, enabling the AI to ask follow-up questions rather than making assumptions when user requests are ambiguous. During training, the model was exposed to both answerable and intentionally unanswerable questions created by removing key pieces of information.
According to Lee, the new model reduced hallucination rates by as much as 30 percentage points compared with the previous HyperCLOVA X model in external benchmark tests.
“Instead of guessing answers to incomplete questions, the model asks for the necessary information, providing users with a more reliable AI experience,” he said.
Naver also introduced harness engineering, a technology that helps ensure AI systems operate safely and effectively in production environments. The framework controls inappropriate responses while enabling the AI to retrieve relevant information, use the right tools, and complete complex user requests.
Han Seung-gyun, AI Search Service Leader at Naver, said the company has implemented multiple safeguards from the service design stage to reduce hallucinations, building on the trust it has established through years of operating its search platform.
AI Tab also employs a collaborative architecture using multiple small language models (SLMs) assigned to specialized roles instead of relying on a single LLM for every task. This approach has reduced hardware operating costs for some components to roughly one-third of previous levels while more than doubling response speed.
Naver also outlined its multimodal AI strategy centered on Smart Lens. Following AI Tab's official launch last month, Smart Lens was prominently positioned next to the AI Tab button, allowing users to identify products through their smartphone cameras and proceed directly to purchases.
The company plans to expand Smart Lens so that a single image input can support not only visual search but also follow-up tasks such as information queries, reservations, and other real-world actions.
Yoon Sang-doo, Head of Naver's Future AI Center, said the company's long-term goal is to develop an action-oriented AI agent capable of understanding what users are looking at, accurately interpreting their intent, connecting with Naver's ecosystem of services, and assisting them in completing tasks.
“We will continue expanding the capabilities of our action-oriented AI agents through multimodal technologies,” Yoon said.
Naver plans to introduce the new capabilities to AI Tab in phases during the second half of the year. This month, the company will integrate its AI Briefing summarization service and Smart Lens into AI Tab. A connection to Naver Real Estate will follow next month, while a health-focused AI agent is scheduled to launch before the end of the year.
