Naver Unveils Optimized AI Tab Model: "Twice the Speed and 30% Reduction in Hallucination Rates"

Lightweight Model Developed for AI Search Service
Reinforcement Learning Significantly Expanded Compared to HyperCLOVA X

Director Ki-chang Lee of Naver Cloud presenting at the 'AI Search Tech Deep Talk' held on the 2nd. Provided by Naver.

Naver has newly integrated a lightweight model into its 'AI Tab' that offers faster response times and higher throughput than the previous hyper-scale artificial intelligence (AI) model, HyperCLOVA X (HCX). Compared to the existing HCX, it increases the use of reinforcement learning, doubles the speed, and improves the hallucination rate by 30 percentage points.

Naver announced these updates on July 2 at the Naver D2SF Gangnam during a Tech Deep Talk session titled "From Search to Execution: How AI Technology Shapes Naver AI Search."

At the session, Naver unveiled key technologies driving next-generation AI search, including the optimized model for conversational AI search 'AI Tab,' harness engineering for safe and efficient AI operation, and multimodal technology that enhances AI’s visual understanding.

Ki-chang Lee, Director of Hyperscale AI Models at Naver Cloud, emphasized, "The model applied to AI Tab is optimized for Naver services across the entire process, from building training data and model design to reinforcement learning." He added, "Our goal is to develop a model that operates most accurately and efficiently in real service scenarios such as search, purchase, and reservation for Naver users."

The newly introduced model on AI Tab is a lightweight version developed for AI search services based on HCX. It is a product-native large language model (LLM) that incorporates Naver’s data, service scenarios, and user feedback into every aspect of model design.

This model was developed with a focus on three pillars—data, architecture, and training—to maximize service efficiency. Naver improved training data quality through document quality filters and built a "service-oriented data collection pipeline" that maps complex user requests to optimal answers. High-quality data from search, shopping, places, and local information was incorporated from the pre-training stage.

In terms of architecture, a Mixture-of-Experts (MoE) structure optimized for large-scale service environments was adopted. This enables faster response times and higher throughput compared to the existing HCX. Notably, the total time from input to final answer completion has been shortened with this model.

During the training phase, computational resources devoted to reinforcement learning more than doubled compared to HCX. By creating a reinforcement learning environment linked to actual services like user simulators and Naver’s search and reservation, the model learned to leverage various tools to complete user tasks to the end.

A new reinforcement learning technique was also introduced, where the model is rewarded for asking follow-up questions when faced with unanswerable queries, thereby improving performance. Clarity reinforcement learning was used to prevent the AI from arbitrarily interpreting ambiguous requests, instead prompting it to clarify user intent through additional questions—reducing hallucinations. According to the AA-Omniscience benchmark by Artificial Analysis, the specialized model’s hallucination rate decreased by up to 30 percentage points compared to the original HCX.

Additionally, the OPD (Own Policy-based Distillation) technique was applied. In this approach, the student model generates answers, which are then refined token-by-token by a high-performance teacher model. This process effectively supplements the student’s weaker areas of expertise, and as the teacher model improves, so does the student, thanks to this enhancement structure.

Naver stated that, according to its own benchmark assessing the execution quality of the model in tasks such as 'search, purchase, and reservation' on AI Tab, it scored 108 points in ‘service capability,’ surpassing the global peer average of 100 points. In terms of basic abilities like instruction following and general tool usage, it scored 104 points, also above the peer average of 100 points.

Lightweight Model on AI Tab Cuts Costs by Up to Threefold

On this day, Naver also unveiled its core technology for reliably operating the agent-based search service AI Tab, called "harness engineering." The harness engineering applied to AI Tab is designed to control the AI so it does not generate inappropriate responses, while ensuring it independently finds necessary information and uses the appropriate tools to fulfill user requests to completion. It operates in four stages: safety filtering, understanding user intent and managing long conversation contexts, service-linked reasoning for search, shopping, and places, and source provision with execution linkage.

To further improve the efficiency of AI Tab, a large-scale service, Naver established a division-of-labor SLM (small language model) structure. Instead of a single large LLM handling all tasks, specialized SLMs are combined by role, reducing operational costs while simultaneously enhancing response speed and quality. In practice, by applying lightweight specialized models to AI Tab, the operational equipment cost for certain components was reduced by up to three times compared to before, and response speed improved by more than twofold. The division-of-labor SLM structure also allows for seamless upgrades with plugin replacements for new small models, enabling continuous performance improvement without service interruptions.

Seungkyun Han, Leader of Naver’s AI Search Service, stressed, "To develop an AI agent that works well in services, harness engineering that designs for both cost efficiency and stability is essential, not just the LLM itself." He added, "The search infrastructure and know-how accumulated over the past 27 years, vast content from blogs and cafes, and diverse assets such as shopping and places, all interconnected by AI technology to deliver an experience from search to execution, are competitive advantages unique to Naver that no one can easily replicate."

Leader Sangdoo Yoon from Naver presenting at the 'AI Search Tech Deep Talk' held on the 2nd. Provided by Naver.

Naver also unveiled its strategy to further advance multimodal technology centered around Smart Lens, which is prominently positioned on the search bar. Multimodal technology enables AI to convert images into representations it can understand, allowing it to process and utilize not only text but also information in images and videos.

Since launching Smart Lens in 2017 and introducing image search services, Naver has continued to advance its technologies, building up its multimodal search capabilities. Sangdoo Yoon, Leader at Naver Future AI Center, stated, "In the future, Naver’s AI agent services will evolve to understand user intent through not just text, but also images, connecting that understanding to real-world actions."

This content was produced with the assistance of AI translation services.