[Dongwon Kim’s Eye-T] From General AI to Expert AI: LG AI Research Institute’s Foresight

A new wave is blowing through the artificial intelligence (AI) industry. Global big tech companies, which had previously focused on creating general-purpose AI that can "do everything," are now actively developing AI specialized in specific fields. Industry-specialized AIs released by big tech include Microsoft’s AI that achieves 85% accuracy in medical diagnosis, Google’s medical AI recognized by 72% of radiologists, and Apple’s AI tailored for personal devices. Interestingly, a Korean institution had already predicted this global trend and acted on it years ago—LG AI Research Institute.

◇ Global Trend: Spotlight on Industry-Specific AI

AI technology is moving out of the lab and creating tangible value in real industrial settings. AI systems that perform at expert levels are now driving actual business growth. The medical field, in particular, is experiencing rapid changes. Google's medical AI model, Med-Gemini, is performing at a specialist level, with 72% of radiologists evaluating its reports as “equal to or better than human doctors.”

Microsoft's diagnostic system MAI-DxO goes even further, achieving 85% accuracy in medical diagnostics. In comparison, the average accuracy among clinicians for the same cases was only 21.2%, showing over four times the performance gap. OpenAI is also setting new benchmarks for healthcare-specialized AI with HealthBench, trained on 262 doctors, 60 countries’ medical practices, and 5,000 real medical dialogues.

Beyond healthcare, industry-specific AIs are emerging in many other fields. In manufacturing, Google developed an AI specialized in semiconductor chip design that reduced work time from months to just six hours. Apple’s Apple Intelligence has pioneered a new direction by focusing on on-device AI for personal devices, prioritizing privacy and contextual understanding, as opposed to cloud-based general AI.

◇ LG AI Research Institute’s Foresight: Specialization from the Start

The result of LG AI Research Institute’s efforts is its large-scale AI model EXAONE, which stands for “Expert AI for EveryONE.” The name reflects its vision of creating expert-level AI for everyone.

In an interview, Choi Jung-kyu, head of the AI Agent Group at LG AI Research Institute, said, “As an independent research organization under LG Group, we tackle complex, high-level problems that are difficult for individual affiliates to solve,” adding, “As we accumulated requests like ‘Can AI solve this?’ we developed large language models (LLMs) capable of understanding and responding precisely to each industry.”

The most notable innovations have come in the chemical and materials sectors. EXAONE Discovery, an AI model based on EXAONE, learned the properties of around 100 types of chemical substances and shortened the core ingredient development process for functional cosmetics from 1 year and 10 months to just one day. LG AI Research Institute unveiled a proprietary technology capable of reading chemical structures, recognizing atoms and bond types within molecules, and converting them into a database. This model showed over 100 times better efficiency than conventional models and was presented at the global AI conference NeurIPS.

In manufacturing, EXAONE is already making practical changes. At LG Display, EXAONE is used in core tasks. The display industry includes a wide variety of protocols and technical documents, ranging from large OLEDs to automotive displays. LG Display had digitized and stored massive R&D issue-resolution documents accumulated over the past 30 years. Based on these, EXAONE is now used as a search-based Q&A service that finds solutions to specific problems from past similar cases—cutting problem-solving time from 20–30 minutes to less than 30 seconds.

In healthcare, LG AI Research Institute is preparing next-generation innovation. Its recently announced EXAONE Path 2.0 is a precision medical AI model that analyzes and predicts genetic mutations, expression patterns, and microstructural features of human tissues and cells using pathology images. It achieved the world’s highest prediction accuracy of 78.4% for genetic mutations. LG AI Research Institute said, “It can reduce the current gene testing time of over two weeks to under one minute, helping secure the golden time for cancer treatment.”

It is also gaining recognition in education. EXAONE 4.0 earned six national professional certifications—including those for doctors, oriental pharmacists, dentists, loss adjusters, appraisers, and customs agents—proving its expert-level learning ability. It even scored 94.5 points in the 2024 Korean CSAT (College Scholastic Ability Test) math section, ranking in the top tier, indicating its potential for use in educational settings.

One representative example is the digital platform project underway with the Gyeonggi Provincial Office of Education and LG CNS. The system supports communication with parents and students at over 280 schools in Gyeonggi Province and provides customized academic schedules and administrative information. LG AI Research Institute plans to offer EXAONE 4.0 for free to K–12 students and university students without a license. It also plans to launch a book search and summary service in cooperation with the Library Association in the second half of the year.

◇ LG AI Research Institute’s Other Strength: AI ‘Fuel Efficiency’ Innovation

A key differentiator for LG AI Research Institute is its “AI fuel efficiency.” It delivers competitive performance with fewer parameters and less training data than global models. It is praised for laying the foundation for reducing costs and power consumption—major pain points in AI deployment—while maintaining high efficiency.

The clearest example is the comparison between EXAONE 3.0 and Meta's LLaMA 3.1. While LLaMA 3.1 was trained on 15 trillion tokens, EXAONE 3.0 used only 8 trillion—nearly half. Given that more data typically means better performance, LLaMA 3.1 should outperform. However, the result was different. EXAONE 3.0 achieved higher scores in actual usability, as well as coding and math benchmarks.

This fuel efficiency is even more pronounced in the latest EXAONE 4.0. Despite having only 32 billion parameters, it achieved top-tier performance comparable to much larger models like DeepSeek-R1 (671B) and Qwen3-235B. Considering that its size is only 4.8% of DeepSeek-R1 and 13.6% of Qwen3-235B, it ranks among the most efficient AI models globally. On the MMLU-Pro benchmark for knowledge and problem-solving, it scored 81.8—close to DeepSeek-R1’s 85 and Qwen3-235B’s 83.

It also shows remarkable performance in on-device models. The lightweight EXAONE Deep-7.8B retains 95% of the 32B model’s performance at just 24% of its size. The 2.4B on-device model achieves 86% performance despite being only 7.5% in size. On-device models process data internally without relying on external servers, offering strong advantages in privacy and security—ideal for deployment in smartphones, automobiles, and robots.

To achieve this fuel efficiency, LG AI Research Institute devoted great effort from pre-training to post-training. It focused on refining raw corpora from the web, books, papers, patents, and code, and eliminating duplicated data. Duplicate data can cause bias by emphasizing repeated content and waste power, reducing efficiency.

Despite aiming for high performance, it did not cut ethical corners. In terms of data use, LG AI Research Institute strictly adhered to ethical standards. At the “THE AI KOREA” conference held in August last year, Jinsik Lee, head of the research lab, said, “There is data that can significantly boost performance, but such data may pose copyright or licensing issues,” and added, “In order to create a safe AI environment, we did not use these data in developing our models, following the LG AI Ethics Principles.” These principles are based on five core values: respect for humanity, fairness, safety, accountability, and transparency.

EXAONE’s global competitiveness has also been proven. Since its open-source release in August 2024, it has been downloaded 3.5 million times worldwide. It was the only Korean model selected by the Stanford Human-Centered AI Institute (HAI) as one of the world’s notable AI models. EXAONE 4.0 showed the highest performance among peer models in reasoning—a domain requiring complex thinking—and has outperformed global competitors in world knowledge, coding, science, and mathematics.

As the paradigm in the global AI market shifts from general-purpose to specialized models, LG AI Research Institute’s EXAONE is considered a milestone pointing the way forward for Korea. On the 22nd, LG AI Research Institute will hold its first “LG AI Talk Concert” in two years to unveil a new innovation.

The key point is that LG is not following the global trend—it is leading it as a true first mover.