Naver has applied next-generation technology to "AI Tab," its conversational artificial intelligence search service. Rather than joining the performance race for general-purpose large language models (LLMs), the company aims to focus on operating quickly and efficiently in real service environments such as search, shopping, and reservations, countering global big tech's push into the domestic search market.
Naver held a press briefing at D2SF in Gangnam, Seoul, on the 2nd, unveiling three core technologies to power its next-generation AI search: product-native LLM, harness engineering, and multimodal technology.
The "product-native LLM" is a lightweight model developed for the AI search service based on HyperCLOVA X, Naver's proprietary LLM. Unlike the ultra-large models of global big tech, it is characterized by a reduced size that concentrates on the performance needed for search.
The transformer structure that underpins existing AI is designed to understand the full context by examining the relationships between words in a sentence one by one. Because of this, as input volume increases, computational load rises by the square, creating a limitation in which response time increases sharply. The product-native LLM, by contrast, changed its structure so that computational load increases only in proportion to how long the question is, so response time does not increase significantly even when processing long documents.
On top of this, "clarity-reinforced learning" was applied, which asks follow-up questions to clarify intent when a user's question is ambiguous. By having the AI first confirm intent instead of arbitrarily fabricating an answer to an unclear question, Naver explained, it reduced the hallucination phenomenon—a chronic problem of generative AI—by 30 percentage points (p) compared with HyperCLOVA X.
The system that supports this LLM in delivering its full performance in actual service is "harness engineering." Naver aims for a "one-stop vertical (specialized field) service" that connects AI search all the way to shopping, reservations, and payment. In this process, harness engineering controls the AI so that it does not produce inappropriate answers, while at the same time grasping user intent and the context of long conversations to smoothly link everything from search to service execution.
Naver also adopted a "division-of-labor SLM (small language model)" structure to improve the efficiency of AI Tab. Instead of a single giant LLM handling all tasks, it combines SLMs specialized by role. Through this, the company said, it lowered equipment operating costs to as much as one-third of the previous level and improved response speed by more than twofold.
It also advanced multimodal technology centered on Smart Lens, its image search function. Multimodal is technology that converts not only text but also images and video into representations (embeddings) that machines can understand. Naver has accumulated multimodal search capabilities since launching Smart Lens in 2017, and this year newly introduced its multimodal LLM "MuCo." The key to MuCo is that it can process an image just once at the start of a conversation and still precisely capture the context of subsequent questions. It resolves the limitations of existing multimodal technology, in which images had to be recomputed each time a question continued, slowing speed and increasing costs. To this end, Naver said it built a multimodal dataset of 35 million and recorded performance exceeding competing models on major multimodal search benchmarks.
Lee Ki-chang, director of hyperscale AI models at Naver Cloud, explained, "Through the models applied to AI Tab, we are providing users with faster and more stable service, and because we can handle more requests with the same graphics processing unit (GPU) resources, we can also significantly reduce operating costs." He added, "Going forward, we plan to continuously create accurate yet light and efficient models."
