%4c%4c%4d
-
Turning PC and Mobile Devices into AI Infrastructure, Reducing ChatGPT Costs
< (From left) KAIST School of Electrical Engineering: Dr. Jinwoo Park, M.S candidate Seunggeun Cho, and Professor Dongsu Han >
Until now, AI services based on Large Language Models (LLMs) have mostly relied on expensive data center GPUs. This has resulted in high operational costs and created a significant barrier to entry for utilizing AI technology. A research team at KAIST has developed a technology that reduces reliance on expensive data center GPUs by utilizing affordable, everyday GPUs to provide AI services at a much lower cost.
On December 28th, KAIST announced that a research team led by Professor Dongsu Han from the School of Electrical Engineering developed 'SpecEdge,' a new technology that significantly lowers LLM infrastructure costs by utilizing affordable, consumer-grade GPUs widely available outside of data centers.
SpecEdge is a system where data center GPUs and "edge GPUs"—found in personal PCs or small servers—collaborate to form an LLM inference infrastructure. By applying this technology, the team successfully reduced the cost per token (the smallest unit of text generated by AI) by approximately 67.6% compared to methods using only data center GPUs.
To achieve this, the research team utilized a method called 'Speculative Decoding.' In this process, a small language model placed on the edge GPU quickly generates a high-probability token sequence (a series of words or word fragments). Then, the large-scale language model in the data center verifies this sequence in batches. During this process, the edge GPU continues to generate words without waiting for the server's response, simultaneously increasing LLM inference speed and infrastructure efficiency.
< Figure 1. Language data flow diagram of the developed SpecEdge >
< Figure 2. Detailed computation time reduction method of SpecEdge >
< Figure 3. Illustration of efficient batching of verification requests from multiple edge GPUs on the server GPU within SpecEdge >
Compared to performing speculative decoding solely on data center GPUs, SpecEdge improved cost efficiency by 1.91 times and server throughput by 2.22 times. Notably, the technology was confirmed to work seamlessly even under standard internet speeds, meaning it can be immediately applied to real-world services without requiring a specialized network environment.
Furthermore, the server is designed to efficiently process verification requests from multiple edge GPUs, allowing it to handle more simultaneous requests without GPU idle time. This has realized an LLM serving infrastructure structure that utilizes data center resources more effectively.
This research presents a new possibility for distributing LLM computations—which were previously concentrated in data centers—to the edge, thereby reducing infrastructure costs and increasing accessibility. In the future, as this expands to various edge devices such as smartphones, personal computers, and Neural Processing Units (NPUs), high-quality AI services are expected to become available to a broader range of users.
< Figure 4. Conceptual comparison of the developed SpecEdge vs. conventional methods >
Professor Dongsu Han, who led the research, stated, "Our goal is to utilize edge resources around the user, beyond the data center, as part of the LLM infrastructure. Through this, we aim to lower AI service costs and create an environment where anyone can utilize high-quality AI."
Dr. Jinwoo Park and M.S candidate Seunggeun Cho from KAIST participated in this study. The research results were presented as a 'Spotlight' (top 3.2% of papers, with a 24.52% acceptance rate) at the NeurIPS (Neural Information Processing Systems) conference, the world's most prestigious academic conference in the field of AI, held in San Diego from December 2nd to 7th.
Paper Title: SpecEdge: Scalable Edge-Assisted Serving Framework for Interactive LLMs
Paper Links: NeurIPS Link, arXiv Link
This research was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) under the project 'Development of 6G System Technology to Support AI-Native Application Services.'
2025.12.29 View 3399 -
KAIST, AI judges manufacturing beyond craftsmanship and language barriers
<(From Left) M.S candidate Inhyo Lee, Ph.D candidate Heekyu Kim, Ph.D candidate joonyoung Kim, Professor Seunghwa Ryu>
Most of the plastic products we use are made through injection molding, a process in which molten plastic is injected into a mold to mass-produce identical items. However, even slight changes in conditions can lead to defects, so the process has long relied on the intuition of highly skilled workers. Now, KAIST researchers have proposed an AI-based solution that autonomously optimizes processes and transfers manufacturing knowledge, addressing concerns that expertise could be lost due to the retirement of skilled workers and the increase in foreign labor.
KAIST (President Kwang Hyung Lee) announced on the 22nd of December that a research team led by Professor Seunghwa Ryu from the Department of Mechanical Engineering · InnoCORE PRISM-AI Center has, for the first time in the world, developed generative AI technology that autonomously optimizes injection molding processes, along with an LLM-based knowledge transfer system that makes on-site expertise accessible to anyone. The team also reported that these achievements were published consecutively in an internationally renowned journal.
The first achievement is a generative AI–based process inference technology that automatically infers optimal process conditions based on environmental changes or quality requirements. Previously, whenever temperature, humidity, or desired quality levels changed, skilled workers had to rely on trial and error to readjust conditions.
The research team implemented a diffusion model–based approach that reverse-engineers process conditions satisfying target quality requirements, using environmental data and process parameters collected over several months from an actual injection molding factory.
In addition, the team built a surrogate model that substitutes for actual production, enabling quality prediction without running the real process. As a result, they achieved an error rate of just 1.63%, significantly lower than the 23~44% error rates of representative existing technologies such as GAN* and VAE** models traditionally used for process prediction. Experiments applying the AI-generated conditions to real processes confirmed successful production of acceptable products, demonstrating practical applicability.
*GAN (Generative Adversarial Network): a method in which two AI models compete with each other to generate data
**VAE (Variational Autoencoder): a method that compresses and learns common patterns in data and then reconstructs them
<Figure 1. Generative AI–Based Process Reasoning Technology>
The second achievement is the IM-Chat, an LLM-based knowledge transfer system designed to address skilled worker retirement and multilingual work environments. IM-Chat is a multi-agent AI system that combines large language models (LLMs) with retrieval-augmented generation (RAG), serving as an AI assistant for manufacturing sites by providing appropriate solutions to problems encountered by novice or foreign workers.
When a worker asks a question in natural language, the AI understands it and, if necessary, automatically calls the generative process inference AI, simultaneously providing optimal process condition calculations along with relevant standards and background explanations.
For example, when asked, “What is the appropriate injection pressure when the factory humidity is 43.5%?”, the AI calculates the optimal condition and presents the supporting manual references as well. With support for multilingual interfaces, foreign workers can receive the same level of decision-making support.
This research is regarded as a core manufacturing AI transformation (AX) technology that can be extended beyond injection molding to molds, presses, extrusion, 3D printing, batteries, bio-manufacturing, and other industries.
In particular, the work is significant in that it presents a paradigm for autonomous manufacturing AI, integrating generative AI and LLM agents through a Tool-Calling approach*, enabling AI to make its own judgments and invoke necessary functions.
*Tool-Calling approach: a method in which AI autonomously calls and uses the functions or programs required for a given situation
<Figure 2. Large Language Model–Based Multilingual Knowledge Transfer Multi-Agent IM-Chat>
<Figure 3. Example of Operation of the Large Language Model (LLM)–Based Multilingual Knowledge Transfer Multi-Agent IM-Chat>
<Figure 4. Illustration of the Application of an LLM-Based Multilingual Knowledge Transfer Multi-Agent IM-Chat (AI-Generated)>
Professor Seunghwa Ryu explained, “This is a case where we addressed fundamental problems in manufacturing in a data-driven way by combining AI that autonomously optimizes processes with LLMs that make on-site knowledge accessible to anyone,” adding, “We will continue expanding this approach to various manufacturing processes to accelerate intelligence and autonomy across the industry.”
This research involved doctoral candidates Junhyeong Lee, Joon-Young Kim, and Heekyu Kim from the Department of Mechanical Engineering as co–first authors, with Professor Seunghwa Ryu as the corresponding author. The results were published consecutively in the April and December issues of Journal of Manufacturing Systems (JCR 1/69, IF 14.2), the world’s top-ranked international journal in engineering and industrial fields.
※ Paper 1: “Development of an Injection Molding Production Condition Inference System Based on Diffusion Model,” DOI: https://doi.org/10.1016/j.jmsy.2025.01.008 ※ Paper 2: “IM-Chat: A multi-agent LLM framework integrating tool-calling and diffusion modeling for knowledge transfer in injection molding industry,” DOI: https://doi.org/10.1016/j.jmsy.2025.11.007
This research was supported by the Ministry of Science and ICT, the Ministry of SMEs and Startups, and the Ministry of Trade, Industry and Energy.
2025.12.22 View 3893 -
KAIST Introduces ‘Virtual Teaching Assistant’ That can Answer Even in the Middle of the Night – Successful First Deployment in Classroom
- Research teams led by Prof. Yoonjae Choi (Kim Jaechul Graduate School of AI) and Prof. Hwajeong Hong (Department of Industrial Design) at KAIST developed a Virtual Teaching Assistant (VTA) to support learning and class operations for a course with 477 students.
- The VTA responds 24/7 to students’ questions related to theory and practice by referencing lecture slides, coding assignments, and lecture videos.
- The system’s source code has been released to support future development of personalized learning support systems and their application in educational settings.
< Photo 1. (From left) PhD candidate Sunjun Kweon, Master's candidate Sooyohn Nam, PhD candidate Hyunseung Lim, Professor Hwajung Hong, Professor Yoonjae Choi >
“At first, I didn’t have high expectations for the Virtual Teaching Assistant (VTA), but it turned out to be extremely helpful—especially when I had sudden questions late at night, I could get immediate answers,” said Jiwon Yang, a Ph.D. student at KAIST. “I was also able to ask questions I would’ve hesitated to bring up with a human TA, which led me to ask even more and ultimately improved my understanding of the course.”
KAIST (President Kwang Hyung Lee) announced on June 5th that a joint research team led by Prof. Yoonjae Choi of the Kim Jaechul Graduate School of AI and Prof. Hwajeong Hong of the Department of Industrial Design has successfully developed and deployed a Virtual Teaching Assistant (VTA) that provides personalized feedback to individual students even in large-scale classes.
This study marks one of the first large-scale, real-world deployments in Korea, where the VTA was introduced in the “Programming for Artificial Intelligence” course at the KAIST Kim Jaechul Graduate School of AI, taken by 477 master’s and Ph.D. students during the Fall 2024 semester, to evaluate its effectiveness and practical applicability in an actual educational setting.
The AI teaching assistant developed in this study is a course-specialized agent, distinct from general-purpose tools like ChatGPT or conventional chatbots. The research team implemented a Retrieval-Augmented Generation (RAG) architecture, which automatically vectorizes a large volume of course materials—including lecture slides, coding assignments, and video lectures—and uses them as the basis for answering students’ questions.
< Photo 2. Teaching Assistant demonstrating to the student how the Virtual Teaching Assistant works>
When a student asks a question, the system searches for the most relevant course materials in real time based on the context of the query, and then generates a response. This process is not merely a simple call to a large language model (LLM), but rather a material-grounded question answering system tailored to the course content—ensuring both high reliability and accuracy in learning support.
Sunjun Kweon, the first author of the study and head teaching assistant for the course, explained, “Previously, TAs were overwhelmed with repetitive and basic questions—such as concepts already covered in class or simple definitions—which made it difficult to focus on more meaningful inquiries.” He added, “After introducing the VTA, students began to reduce repeated questions and focus on more essential ones. As a result, the burden on TAs was significantly reduced, allowing us to concentrate on providing more advanced learning support.”
In fact, compared to the previous year’s course, the number of questions that required direct responses from human TAs decreased by approximately 40%.
< Photo 3. A student working with VTA. >
The VTA, which was operated over a 14-week period, was actively used by more than half of the enrolled students, with a total of 3,869 Q&A interactions recorded. Notably, students without a background in AI or with limited prior knowledge tended to use the VTA more frequently, indicating that the system provided practical support as a learning aid, especially for those who needed it most.
The analysis also showed that students tended to ask the VTA more frequently about theoretical concepts than they did with human TAs. This suggests that the AI teaching assistant created an environment where students felt free to ask questions without fear of judgment or discomfort, thereby encouraging more active engagement in the learning process.
According to surveys conducted before, during, and after the course, students reported increased trust, response relevance, and comfort with the VTA over time. In particular, students who had previously hesitated to ask human TAs questions showed higher levels of satisfaction when interacting with the AI teaching assistant.
< Figure 1. Internal structure of the AI Teaching Assistant (VTA) applied in this course. It follows a Retrieval-Augmented Generation (RAG) structure that builds a vector database from course materials (PDFs, recorded lectures, coding practice materials, etc.), searches for relevant documents based on student questions and conversation history, and then generates responses based on them. >
Professor Yoonjae Choi, the lead instructor of the course and principal investigator of the study, stated, “The significance of this research lies in demonstrating that AI technology can provide practical support to both students and instructors. We hope to see this technology expanded to a wider range of courses in the future.”
The research team has released the system’s source code on GitHub, enabling other educational institutions and researchers to develop their own customized learning support systems and apply them in real-world classroom settings.
< Figure 2. Initial screen of the AI Teaching Assistant (VTA) introduced in the "Programming for AI" course. It asks for student ID input along with simple guidelines, a mechanism to ensure that only registered students can use it, blocking indiscriminate external access and ensuring limited use based on students. >
The related paper, titled “A Large-Scale Real-World Evaluation of an LLM-Based Virtual Teaching Assistant,” was accepted on May 9, 2025, to the Industry Track of ACL 2025, one of the most prestigious international conferences in the field of Natural Language Processing (NLP), recognizing the excellence of the research.
< Figure 3. Example conversation with the AI Teaching Assistant (VTA). When a student inputs a class-related question, the system internally searches for relevant class materials and then generates an answer based on them. In this way, VTA provides learning support by reflecting class content in context. >
This research was conducted with the support of the KAIST Center for Teaching and Learning Innovation, the National Research Foundation of Korea, and the National IT Industry Promotion Agency.
2025.06.05 View 7678 -
“For the First Time, We Shared a Meaningful Exchange”: KAIST Develops an AI App for Parents and Minimally Verbal Autistic Children Connect
• KAIST team up with NAVER AI Lab and Dodakim Child Development Center Develop ‘AAcessTalk’, an AI-driven Communication Tool bridging the gap Between Children with Autism and their Parents
• The project earned the prestigious Best Paper Award at the ACM CHI 2025, the Premier International Conference in Human-Computer Interaction
• Families share heartwarming stories of breakthrough communication and newfound understanding.
< Photo 1. (From left) Professor Hwajung Hong and Doctoral candidate Dasom Choi of the Department of Industrial Design with SoHyun Park and Young-Ho Kim of Naver Cloud AI Lab >
For many families of minimally verbal autistic (MVA) children, communication often feels like an uphill battle. But now, thanks to a new AI-powered app developed by researchers at KAIST in collaboration with NAVER AI Lab and Dodakim Child Development Center, parents are finally experiencing moments of genuine connection with their children.
On the 16th, the KAIST (President Kwang Hyung Lee) research team, led by Professor Hwajung Hong of the Department of Industrial Design, announced the development of ‘AAcessTalk,’ an artificial intelligence (AI)-based communication tool that enables genuine communication between children with autism and their parents.
This research was recognized for its human-centered AI approach and received international attention, earning the Best Paper Award at the ACM CHI 2025*, an international conference held in Yokohama, Japan.*ACM CHI (ACM Conference on Human Factors in Computing Systems) 2025: One of the world's most prestigious academic conference in the field of Human-Computer Interaction (HCI).
This year, approximately 1,200 papers were selected out of about 5,000 submissions, with the Best Paper Award given to only the top 1%. The conference, which drew over 5,000 researchers, was the largest in its history, reflecting the growing interest in ‘Human-AI Interaction.’
Called AACessTalk, the app offers personalized vocabulary cards tailored to each child’s interests and context, while guiding parents through conversations with customized prompts. This creates a space where children’s voices can finally be heard—and where parents and children can connect on a deeper level.
Traditional augmentative and alternative communication (AAC) tools have relied heavily on fixed card systems that often fail to capture the subtle emotions and shifting interests of children with autism. AACessTalk breaks new ground by integrating AI technology that adapts in real time to the child’s mood and environment.
< Figure. Schematics of AACessTalk system. It provides personalized vocabulary cards for children with autism and context-based conversation guides for parents to focus on practical communication. Large ‘Turn Pass Button’ is placed at the child’s side to allow the child to lead the conversation. >
Among its standout features is a large ‘Turn Pass Button’ that gives children control over when to start or end conversations—allowing them to lead with agency. Another feature, the “What about Mom/Dad?” button, encourages children to ask about their parents’ thoughts, fostering mutual engagement in dialogue, something many children had never done before.
One parent shared, “For the first time, we shared a meaningful exchange.” Such stories were common among the 11 families who participated in a two-week pilot study, where children used the app to take more initiative in conversations and parents discovered new layers of their children’s language abilities.
Parents also reported moments of surprise and joy when their children used unexpected words or took the lead in conversations, breaking free from repetitive patterns. “I was amazed when my child used a word I hadn’t heard before. It helped me understand them in a whole new way,” recalled one caregiver.
Professor Hwajung Hong, who led the research at KAIST’s Department of Industrial Design, emphasized the importance of empowering children to express their own voices. “This study shows that AI can be more than a communication aid—it can be a bridge to genuine connection and understanding within families,” she said.
Looking ahead, the team plans to refine and expand human-centered AI technologies that honor neurodiversity, with a focus on bringing practical solutions to socially vulnerable groups and enriching user experiences.
This research is the result of KAIST Department of Industrial Design doctoral student Dasom Choi's internship at NAVER AI Lab.* Thesis Title: AACessTalk: Fostering Communication between Minimally Verbal Autistic Children and Parents with Contextual Guidance and Card Recommendation* DOI: 10.1145/3706598.3713792* Main Author Information: Dasom Choi (KAIST, NAVER AI Lab, First Author), SoHyun Park (NAVER AI Lab) , Kyungah Lee (Dodakim Child Development Center), Hwajung Hong (KAIST), and Young-Ho Kim (NAVER AI Lab, Corresponding Author)
This research was supported by the NAVER AI Lab internship program and grants from the National Research Foundation of Korea: the Doctoral Student Research Encouragement Grant (NRF-2024S1A5B5A19043580) and the Mid-Career Researcher Support Program for the Development of a Generative AI-Based Augmentative and Alternative Communication System for Autism Spectrum Disorder (RS-2024-00458557).
2025.05.19 View 13976 -
KAIST to Join Forces with Northwestern School of Law in MIP Program
KAIST has agreed to collaborate in its Master of Intellectual Property (MIP) program with Northwestern University School of Law, the university authorities said on Wednesday (Nov. 11).
KAIST President Nam-Pyo Suh and Northwestern Univ. School of Law Dean David Van Zandt signed a memorandum of understanding for the establishment of the certificate and degree programs of the Anglo-American law in February 2010, at the president"s office Wednesday.
The latest agreement represents an expansion of the existing KAIST-Northwestern Executive LLM program that the two schools have successfully operated for the past eight years. It is aimed at boosting KAIST"s global intellectual property capabilities at a time when the strategic management of intellectual properties and capabilities to deal with international intellectual property disputes are gaining growing importance.
The newly-established two-year, six-semester program is designed to offer a certificate program in the Anglo-American and intellectual property laws to make students better armed with legal knowledge in a competitive global environment.
Northwestern University School of Law is regarded as one of the top law schools in the United States. American law schools are better positioned than any other institutions to prepare students to move from domestic to international practice in today"s complex global economy.
LLM is Latin for Legum Magister, signifying Master of Laws
. course and neto ba ... signing an agreement to run a joint masters degree program at the KAIST campus in Seou
2009.11.19 View 21109