Top Scores on Korean Language Expert-Level Evaluation
Multimodal Integration... Visual Reasoning Technology Secured
Plans to Release Reasoning Model as Open Source
Naver announced on June 30 that it has completed the development of its proprietary generative artificial intelligence (AI) model, 'HyperCLOVA X THINK,' which enhances reasoning capabilities. On the same day, Naver released a technical report detailing the model's architecture, performance, and other key information.
The reasoning model is an AI with enhanced "thinking power." When a user enters a query, the model responds by engaging in extended internal reasoning, almost as if thinking aloud. In this process, the model demonstrates the ability to break down complex problems into smaller components, select appropriate tools or functions, and reflect on and correct its own mistakes, thereby increasing the accuracy and usefulness of the information it provides.
Naver has completed the development of the generative artificial intelligence (AI) 'HyperCLOVA X THINK,' which enhances reasoning capabilities, and has released a technical report. Provided by Naver
Based on its reasoning capabilities, HyperCLOVA X THINK has achieved a high level of language understanding. According to Naver, when evaluated using the 'KoBALT-700' benchmark, which measures the language proficiency of major large language models (LLMs), HyperCLOVA X THINK scored higher than other domestic reasoning models of similar scale and leading global open-source models.
This benchmark was designed by the Department of Linguistics at Seoul National University to assess the Korean language comprehension of LLMs. It consists of expert-level questions evaluating whether the AI accurately understands conversational maxims and analyzes the argument structure of sentences.
HyperCLOVA X THINK also achieved higher scores than major domestic and international open-source models, including those with reasoning capabilities, on another representative Korean language performance metric, the 'HAERAE-Bench.'
Furthermore, Naver has equipped HyperCLOVA X THINK with the ability to reason not only based on language but also on visual information. According to the technical report, HyperCLOVA X THINK was able to recognize and reason through STEM (Science, Technology, Engineering, Mathematics) problems presented in image format and arrive at the correct answers.
For example, in a college entrance exam biology question, the model recognized and analyzed illustrations depicting the "process of ecological succession" and "graphs of total productivity and respiration of a specific plant community over time." It then combined this analysis with relevant knowledge to select the correct statement from multiple choices.
Yoo Kangmin, leader of Naver Cloud's performance evaluation team, stated, "Although this reasoning model was not specifically designed for multimodal reasoning, it has produced meaningful results in the area of visual reasoning. Since we have already secured image, video, and audio multimodal technologies based on HyperCLOVA X, we will further advance the model to achieve even more powerful multimodal reasoning capabilities in the future."
Naver plans to release the reasoning model as open source. Previously, the open-source lightweight model 'HyperCLOVA X SEED,' released by Naver in April, surpassed 500,000 downloads within just over a month.
Sung Nakho, head of hyperscale AI technology at Naver Cloud, said, "We are advancing HyperCLOVA X along two axes: 'enhancement of intelligence' and 'expansion of perception.' With HyperCLOVA X THINK, we have made significant progress in terms of intelligence. We will continue to seek ways to provide users with tangible value, not just keeping pace with technological paradigms."
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
