%61%69%20%6d%6f%64%65%6c
-
How Does AI Think? KAIST Achieves First Visualization of the Internal Structure Behind AI Decision-Making
<(From Left) Ph.D candidate Daehee Kwon, Ph.D candidate Sehyun lee, Professor Jaesik Choi>
Although deep learning–based image recognition technology is rapidly advancing, it still remains difficult to clearly explain the criteria AI uses internally to observe and judge images. In particular, technologies that analyze how large-scale models combine various concepts (e.g., cat ears, car wheels) to reach a conclusion have long been recognized as a major unsolved challenge.
KAIST (President Kwang Hyung Lee) announced on the 26th of November that Professor Jaesik Choi’s research team at the Kim Jaechul Graduate School of AI has developed a new explainable AI (XAI) technology that visualizes the concept-formation process inside a model at the level of circuits, enabling humans to understand the basis on which AI makes decisions.
The study is evaluated as a significant step forward that allows researchers to structurally examine “how AI thinks.”
Inside deep learning models, there exist basic computational units called neurons, which function similarly to those in the human brain. Neurons detect small features within an image—such as the shape of an ear, a specific color, or an outline—and compute a value (signal) that is transmitted to the next layer.
In contrast, a circuit refers to a structure in which multiple neurons are connected to jointly recognize a single meaning (concept). For example, to recognize the concept of cat ear, neurons detecting outline shapes, neurons detecting triangular forms, and neurons detecting fur-color patterns must activate in sequence, forming a functional unit (circuit).
Up until now, most explanation techniques have taken a neuron-centric approach based on the idea that “a specific neuron detects a specific concept.” However, in reality, deep learning models form concepts through cooperative circuit structures involving many neurons. Based on this observation, the KAIST research team proposed a technique that expands the unit of concept representation from “neuron → circuit.”
The research team’s newly developed technology, Granular Concept Circuits (GCC), is a novel method that analyzes and visualizes how an image-classification model internally forms concepts at the circuit level.
GCC automatically traces circuits by computing Neuron Sensitivity and Semantic Flow. Neuron Sensitivity indicates how strongly a neuron responds to a particular feature, while Semantic Flow measures how strongly that feature is passed on to the next concept. Using these metrics, the system can visualize, step-by-step, how basic features such as color and texture are assembled into higher-level concepts.
The team conducted experiments in which specific circuits were temporarily disabled (ablation). As a result, when the circuit responsible for a concept was deactivated, the AI’s predictions actually changed.
In other words, the experiment directly demonstrated that the corresponding circuit indeed performs the function of recognizing that concept.
This study is regarded as the first to reveal, at a fine-grained circuit level, the actual structural process by which concepts are formed inside complex deep learning models. Through this, the research suggests practical applicability across the entire explainable AI (XAI) domain—including strengthening transparency in AI decision-making, analyzing the causes of misclassification, detecting bias, improving model debugging and architecture, and enhancing safety and accountability.
The research team stated, “This technology shows the concept structures that AI forms internally in a way that humans can understand,” adding that “this study provides a scientific starting point for researching how AI thinks.”
Professor Jaesik Choi emphasized, “Unlike previous approaches that simplified complex models for explanation, this is the first approach to precisely interpret the model’s interior at the level of fine-grained circuits,” and added, “We demonstrated that the concepts learned by AI can be automatically traced and visualized.”
< Overview of the Conceptual Circuit Proposed by the Research Team >
This study, with Ph.D. candidates Dahee Kwon and Sehyun Lee from KAIST Kim Jaechul Graduate School of AI as co–first authors, was presented on October 21 at the International Conference on Computer Vision (ICCV).
Paper title: Granular Concept Circuits: Toward a Fine-Grained Circuit Discovery for Concept Representations
Paper link: https://openaccess.thecvf.com/content/ICCV2025/papers/Kwon_Granular_Concept_Circuits_Toward_a_Fine-Grained_Circuit_Discovery_for_Concept_ICCV_2025_paper.pdf
This research was supported by the Ministry of Science and ICT and the Institute for Information & Communications Technology Planning & Evaluation (IITP) under the “Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation” project, the AI Research Hub Project, and the KAIST AI Graduate School Program, and was carried out with support from the Defense Acquisition Program Administration (DAPA) and the Agency for Defense Development (ADD) at the KAIST Center for Applied Research in Artificial Intelligence.
2025.11.26 View 3225 -
KAIST Develops AI That Automatically Designs Optimal Drug Candidates for Cancer-Targeting Mutations
< (From left) Ph.D candidate Wonho Zhung, Ph.D cadidate Joongwon Lee , Prof. Woo Young Kim , Ph.D candidate Jisu Seo >
Traditional drug development methods involve identifying a target protin (e.g., a cancer cell receptor) that causes disease, and then searching through countless molecular candidates (potential drugs) that could bind to that protein and block its function. This process is costly, time-consuming, and has a low success rate. KAIST researchers have developed an AI model that, using only information about the target protein, can design optimal drug candidates without any prior molecular data—opening up new possibilities for drug discovery.
KAIST (President Kwang Hyung Lee) announced on the 10th that a research team led by Professor Woo Youn Kim in the Department of Chemistry has developed an AI model named BInD (Bond and Interaction-generating Diffusion model), which can design and optimize drug candidate molecules tailored to a protein’s structure alone—without needing prior information about binding molecules. The model also predicts the binding mechanism (non-covalent interactions) between the drug and the target protein.
The core innovation of this technology lies in its “simultaneous design” approach. Previous AI models either focused on generating molecules or separately evaluating whether the generated molecule could bind to the target protein. In contrast, this new model considers the binding mechanism between the molecule and the protein during the generation process, enabling comprehensive design in one step. Since it pre-accounts for critical factors in protein-ligand binding, it has a much higher likelihood of generating effective and stable molecules. The generation process visually demonstrates how types and positions of atoms, covalent bonds, and interactions are created simultaneously to fit the protein’s binding site.
<Figure 1. Schematic of the diffusion model developed by the research team, which generates molecular structures and non-covalent interactions based on protein structures. Starting from a noise distribution, the model gradually removes noise (via reverse diffusion) to restore the atom positions, types, covalent bond types, and interaction types, thereby generating molecules. Interacting patterns are extracted from prior knowledge of known binding molecules or proteins, and through an inpainting technique, these patterns are kept fixed during the reverse diffusion process to guide the molecular generation.>
Moreover, this model is designed to meet multiple essential drug design criteria simultaneously—such as target binding affinity, drug-like properties, and structural stability. Traditional models often optimized for only one or two goals at the expense of others, but this new model balances various objectives, significantly enhancing its practical applicability.
The research team explained that the AI operates based on a “diffusion model”—a generative approach where a structure becomes increasingly refined from a random state. This is the same type of model used in AlphaFold 3, the 2024 Nobel Chemistry Prize-winning tool for protein-ligand structure generation, which has already demonstrated high efficiency.
Unlike AlphaFold 3, which provides spatial coordinates for atom positions, this study introduced a knowledge-based guide grounded in actual chemical laws—such as bond lengths and protein-ligand distances—enabling more chemically realistic structure generation.
<Figure 2. (Left) Target protein and the original bound molecule; (Right) Examples of molecules designed using the model developed in this study. The values for protein binding affinity (Vina), drug-likeness (QED), and synthetic accessibility (SA) are shown at the bottom.>
Additionally, the team applied an optimization strategy where outstanding binding patterns from prior results are reused. This allowed the model to generate even better drug candidates without additional training. Notably, the AI successfully produced molecules that selectively bind to the mutated residues of EGFR, a cancer-related target protein.
This study is also meaningful because it advances beyond the team’s previous research, which required prior input about the molecular conditions for the interaction pattern of protein binding.
Professor Woo Youn Kim commented that “the newly developed AI can learn and understand the key features required for strong binding to a target protein, and design optimal drug candidate molecules—even without any prior input. This could significantly shift the paradigm of drug development.” He added, “Since this technology generates molecular structures based on principles of chemical interactions, it is expected to enable faster and more reliable drug development.”
Joongwon Lee and Wonho Zhung, PhD students in the Department of Chemistry, participated as co-first authors of this study. The research results were published in the international journal Advanced Science (IF = 14.1) on July 11.
● Paper Title: BInD: Bond and Interaction-Generating Diffusion Model for Multi-Objective Structure-Based Drug Design
● DOI: 10.1002/advs.202502702
This research was supported by the National Research Foundation of Korea and the Ministry of Health and Welfare.
2025.08.12 View 5524 -
Material Innovation Realized with Robotic Arms and AI, Without Human Researchers
<(From Left) M.S candidate Dongwoo Kim from KAIST, Ph.D candidate Hyun-Gi Lee from KAIST, Intern Yeham Kang from KAIST, M.S candidate Seongjae Bae from KAIST, Professor Dong-Hwa Seo from KAIST, (From top right, from left) Senior Researcher Inchul Park from POSCO Holdings, Senior Researcher Jung Woo Park, senior researcher from POSCO Holdings>
A joint research team from industry and academia in Korea has successfully developed an autonomous lab that uses AI and automation to create new cathode materials for secondary batteries. This system operates without human intervention, drastically reducing researcher labor and cutting the material discovery period by 93%.
* Autonomous Lab: A platform that autonomously designs, conducts, and analyzes experiments to find the optimal material.
KAIST (President Kwang Hyung Lee) announced on the 3rd of August that the research team led by Professor Dong-Hwa Seo of the Department of Materials Science and Engineering, in collaboration with the team of LIB Materials Research Center in Energy Materials R&D Laboratories at POSCO Holdings' POSCO N.EX.T Hub (Director Ki Soo Kim), built the lab to explore cathode materials using AI and automation technology.
Developing secondary battery cathode materials is a labor-intensive and time-consuming process for skilled researchers. It involves extensive exploration of various compositions and experimental variables through weighing, transporting, mixing, sintering*, and analyzing samples.
* Sintering: A process in which powder particles are heated to form a single solid mass through thermal activation.
The research team's autonomous lab combines an automated system with an AI model. The system handles all experimental steps—weighing, mixing, pelletizing, sintering, and analysis—without human interference. The AI model then interprets the data, learns from it, and selects the best candidates for the next experiment.
<Figure 1. Outline of the Anode Material Autonomous Exploration Laboratory>
To increase efficiency, the team designed the automation system with separate modules for each process, which are managed by a central robotic arm. This modular approach reduces the system's reliance on the robotic arm.
The team also significantly improved the synthesis speed by using a new high-speed sintering method, which is 50 times faster than the conventional low-speed method. This allows the autonomous lab to acquire 12 times more material data compared to traditional, researcher-led experiments.
<Figure 2. Synthesis of Cathode Material Using a High-Speed Sintering Device>
The vast amount of data collected is automatically interpreted by the AI model to extract information such as synthesized phases and impurity ratios. This data is systematically stored to create a high-quality database, which then serves as training data for an optimization AI model. This creates a closed-loop experimental system that recommends the next cathode composition and synthesis conditions for the automated system.
* Closed-loop experimental system: A system that independently performs all experimental processes without researcher intervention.
Operating this intelligent automation system 24 hours a day can secure more than 12 times the experimental data and shorten material discovery time by 93%. For a project requiring 500 experiments, the system can complete the work in about 6 days, whereas a traditional researcher-led approach would take 84 days.
During development, POSCO Holdings team managed the overall project planning, reviewed the platform design, and co-developed the partial module design and AI-based experimental model. The KAIST team, led by Professor Dong-hwa Seo, was responsible for the actual system implementation and operation, including platform design, module fabrication, algorithm creation, and system verification and improvement.
Professor Dong-Hwa Seo of KAIST stated that this system is a solution to the decrease in research personnel due to the low birth rate in Korea. He expects it will enhance global competitiveness by accelerating secondary battery material development through the acquisition of high-quality data.
<Figure 3. Exterior View (Side) of the Cathode Material Autonomous Exploration Laboratory>
POSCO N.EX.T Hub plans to apply an upgraded version of this autonomous lab to its own research facilities after 2026 to dramatically speed up next-generation secondary battery material development. They are planning further developments to enhance the system's stability and scalability, and hope this industry-academia collaboration will serve as a model for using innovative technology in real-world R&D.
<Figure 4. Exterior View (Front) of the Cathode Material Autonomous Exploration Laboratory>
The research was spearheaded by Ph.D. student Hyun-Gi Lee, along with master's students Seongjae Bae and Dongwoo Kim from Professor Dong-Hwa Seo’s lab at KAIST. Senior researchers Jung Woo Park and Inchul Park from LIB Materials Research Center of POSCO N.EX.T Hub's Energy Materials R&D Laboratories (Director Jeongjin Hong) also participated.
2025.08.06 View 5907 -
KAIST researcher Se Jin Park develops 'SpeechSSM,' opening up possibilities for a 24-hour AI voice assistant.
<(From Left)Prof. Yong Man Ro and Ph.D. candidate Sejin Park>
Se Jin Park, a researcher from Professor Yong Man Ro’s team at KAIST, has announced 'SpeechSSM', a spoken language model capable of generating long-duration speech that sounds natural and remains consistent.
An efficient processing technique based on linear sequence modeling overcomes the limitations of existing spoken language models, enabling high-quality speech generation without time constraints.
It is expected to be widely used in podcasts, audiobooks, and voice assistants due to its ability to generate natural, long-duration speech like humans.
Recently, Spoken Language Models (SLMs) have been spotlighted as next-generation technology that surpasses the limitations of text-based language models by learning human speech without text to understand and generate linguistic and non-linguistic information. However, existing models showed significant limitations in generating long-duration content required for podcasts, audiobooks, and voice assistants. Now, KAIST researcher has succeeded in overcoming these limitations by developing 'SpeechSSM,' which enables consistent and natural speech generation without time constraints.
KAIST(President Kwang Hyung Lee) announced on the 3rd of July that Ph.D. candidate Sejin Park from Professor Yong Man Ro's research team in the School of Electrical Engineering has developed 'SpeechSSM,' a spoken. a spoken language model capable of generating long-duration speech.
This research is set to be presented as an oral paper at ICML (International Conference on Machine Learning) 2025, one of the top machine learning conferences, selected among approximately 1% of all submitted papers. This not only proves outstanding research ability but also serves as an opportunity to once again demonstrate KAIST's world-leading AI research capabilities.
A major advantage of Spoken Language Models (SLMs) is their ability to directly process speech without intermediate text conversion, leveraging the unique acoustic characteristics of human speakers, allowing for the rapid generation of high-quality speech even in large-scale models.
However, existing models faced difficulties in maintaining semantic and speaker consistency for long-duration speech due to increased 'speech token resolution' and memory consumption when capturing very detailed information by breaking down speech into fine fragments.
To solve this problem, Se Jin Park developed 'SpeechSSM,' a spoken language model using a Hybrid State-Space Model, designed to efficiently process and generate long speech sequences.
This model employs a 'hybrid structure' that alternately places 'attention layers' focusing on recent information and 'recurrent layers' that remember the overall narrative flow (long-term context). This allows the story to flow smoothly without losing coherence even when generating speech for a long time. Furthermore, memory usage and computational load do not increase sharply with input length, enabling stable and efficient learning and the generation of long-duration speech.
SpeechSSM effectively processes unbounded speech sequences by dividing speech data into short, fixed units (windows), processing each unit independently, and then combining them to create long speech.
Additionally, in the speech generation phase, it uses a 'Non-Autoregressive' audio synthesis model (SoundStorm), which rapidly generates multiple parts at once instead of slowly creating one character or one word at a time, enabling the fast generation of high-quality speech.
While existing models typically evaluated short speech models of about 10 seconds, Se Jin Park created new evaluation tasks for speech generation based on their self-built benchmark dataset, 'LibriSpeech-Long,' capable of generating up to 16 minutes of speech.
Compared to PPL (Perplexity), an existing speech model evaluation metric that only indicates grammatical correctness, she proposed new evaluation metrics such as 'SC-L (semantic coherence over time)' to assess content coherence over time, and 'N-MOS-T (naturalness mean opinion score over time)' to evaluate naturalness over time, enabling more effective and precise evaluation.
Through these new evaluations, it was confirmed that speech generated by the SpeechSSM spoken language model consistently featured specific individuals mentioned in the initial prompt, and new characters and events unfolded naturally and contextually consistently, despite long-duration generation. This contrasts sharply with existing models, which tended to easily lose their topic and exhibit repetition during long-duration generation.
PhD candidate Sejin Park explained, "Existing spoken language models had limitations in long-duration generation, so our goal was to develop a spoken language model capable of generating long-duration speech for actual human use." She added, "This research achievement is expected to greatly contribute to various types of voice content creation and voice AI fields like voice assistants, by maintaining consistent content in long contexts and responding more efficiently and quickly in real time than existing methods."
This research, with Se Jin Park as the first author, was conducted in collaboration with Google DeepMind and is scheduled to be presented as an oral presentation at ICML (International Conference on Machine Learning) 2025 on July 16th.
Paper Title: Long-Form Speech Generation with Spoken Language Models
DOI: 10.48550/arXiv.2412.18603
Ph.D. candidate Se Jin Park has demonstrated outstanding research capabilities as a member of Professor Yong Man Ro's MLLM (multimodal large language model) research team, through her work integrating vision, speech, and language. Her achievements include a spotlight paper presentation at 2024 CVPR (Computer Vision and Pattern Recognition) and an Outstanding Paper Award at 2024 ACL (Association for Computational Linguistics).
For more information, you can refer to the publication and accompanying demo: SpeechSSM Publications.
2025.07.04 View 4676 -
King Saud University and KAIST discussed Strategic AI Partnership
<From left> President Abdulla Al-Salman(King Saud University), President Kwang Hyung Lee(KAIST)
KAIST (President Kwang Hyung Lee) and King Saud University (President Abdulla Al-Salman) held a meeting on July 3 at the KAIST Campus in Seoul and agreed to pursue strategic cooperation in AI and digital platform development. The global AI landscape is increasingly polarized between closed models developed by the U.S. and China’s nationally focused technology ecosystems. In this context, many neutral countries have consistently called for an alternative third model that promotes both technological diversity and open access. President Lee has previously advocated for a "Tripartite Platform Strategy" (三分之計), proposing an international collaboration framework based on open-source principles to be free from binary digital power structures and foster cooperative coexistence.
This KAIST-KSU collaboration represents a step toward developing a new, inclusive AI model. The collaboration aims to establish an innovative multilateral framework, especially within the MENA, Japan, Korea, and Southeast Asia, by building an open-source-based AI alliance. Both institutions bring complementary strengths to the table. Saudi Arabia possesses large-scale capital and digital infrastructure, while Korea leads in core AI and semiconductor technologies, applied research, and talent cultivation.
Together, the two nations aim to establish a sustainable collaboration model that creates a virtuous cycle of investment, technology, and talent. This initiative is expected to contribute to the development of an open AI platform and promote diversity in the global AI ecosystem.
During the meeting, the two sides discussed key areas of future cooperation, including:
· Joint development of open-source AI technologies and digital platforms
· Launch of a KAIST-KSU dual graduate degree program
· Expansion of exchange programs for students, faculty, and researchers
· Collaborative research in basic science and STEM disciplines
In particular, the two institutions discussed to establish a joint AI research center to co-develop open AI models and explore practical industrial applications. The goal is to broaden access to AI technology and create an inclusive innovation environment for more countries and institutions.
President Abdulla Al-Salman stated, "Under Saudi Vision 2030, we are driving innovation in science and technology through new leadership, openness, and strategic investment. This partnership with KAIST will serve as a critical foundation for building a competitive AI ecosystem in the Middle East."
President Kwang Hyung Lee emphasized, "By combining Saudi Arabia's leadership, market, and investment capacity with KAIST's technological innovation and the rich talent pools from both countries, we will significantly contribute to diversifying the global AI ecosystem."
Both leaders further noted, "Through joint research leading to an independent AI model, our two institutions could establish a new axis beyond the existing US-China digital order—realizing a 'Tripartite AI Strategy' that will propel us into global markets extending far beyond the MENA and ASEAN regions."
KAIST and KSU plan to formalize this agreement by signing an MOU in the near future, followed by concrete actions such as launching the joint research institute and global talent development programs. This collaboration was initiated under the Korea Foundation’s Distinguished Guests Invitation Program, overseen by the Ministry of Foreign Affairs, and is expected to grow into a long-term strategic partnership with continued support from KF.
About King Saud University (KSU)
Founded in 1957, KSU is Saudi Arabia’s first and leading national university. As a top research-oriented institution in the Middle East, it has achieved international recognition in fields such as AI, energy, and biotechnology. It plays a central role in nurturing talent and driving innovation aligned with Saudi Arabia’s Vision 2030, and is expanding global partnerships to further strengthen its research capabilities.
About the Korea Foundation (KF)
Established in 1991 under the Ministry of Foreign Affairs, the Korea Foundation is a public diplomacy institution dedicated to strengthening international understanding and friendship with Korea. KF plays a key role in expanding Korea’s soft power through academic and cultural exchange, people-to-people networks, and global Korean studies programs. Its Distinguished Guests Invitation Program fosters strategic partnerships with global leaders in government, academia, and industry.
2025.07.04 View 5165 -
Opening the Door to B Cell-Based Cancer Remembering Personalized Cancer Vaccines
< (From left) KAIST Professor Jung Kyoon Choi, Dr. Jeong Yeon Kim, and Dr. Jin Hyeon An >
Neoantigens are unique markers that distinguish only cancer cells. By adding B cell reactivity, cancer vaccines can move beyond one-time attacks and short-term memory to become a long-term immunity that "remembers" cancer, effectively preventing recurrence. KAIST’s research team has developed an AI-based personalized cancer vaccine design technology that makes this possible and optimizes anticancer effects for each individual.
KAIST announced on January 2nd that Professor Jung Kyoon Choi’s research team from the Department of Bio and Brain Engineering, in collaboration with Neogen Logic Co., Ltd., has developed a new AI model to predict neoantigens—a core element of personalized cancer vaccine development—and clarified the importance of B cells in cancer immunotherapy.
The research team overcame the limitations of existing neoantigen discovery, which relied primarily on predicting T cell reactivity, and developed an AI-based neoantigen prediction technology that integrally considers both T cell and B cell reactivity.
This technology has been validated through large-scale cancer genome data, animal experiments, and clinical trial data for cancer vaccines. It is evaluated as the first AI technology capable of quantitatively predicting B cell reactivity to neoantigens.
Neoantigens are antigens composed of protein fragments derived from cancer cell mutations. Because they possess cancer-cell specificity, they have gained attention as a core target for next-generation cancer vaccines. Companies like Moderna and BioNTech developed COVID-19 vaccines using the mRNA platforms they secured while advancing neoantigen-based cancer vaccine technology, and they are currently actively conducting clinical trials for cancer vaccines alongside global pharmaceutical companies.
However, current cancer vaccine technology is mostly focused on T cell-centered immune responses, presenting a limitation in that it does not sufficiently reflect the immune responses mediated by B cells.
In fact, the research team of Professors Mark Yarchoan and Elizabeth Jaffee at Johns Hopkins University pointed out in Nature Reviews Cancer in May 2025 that “despite accumulating evidence regarding the role of B cells in tumor immunity, most cancer vaccine clinical trials still focus only on T cell responses.”
The research team’s new AI model overcomes existing limitations by learning the structural binding characteristics between mutant proteins and B cell receptors (BCR) to predict B cell reactivity. In particular, an analysis of cancer vaccine clinical trial data confirmed that integrating B cell responses can significantly enhance anti-tumor immune effects in actual clinical settings.
< Schematic Background of the Technology >
Professor Jung Kyoon Choi stated, “Together with Neogen Logic Co., Ltd., which is currently commercializing neoantigen AI technology, we are conducting pre-clinical development of a personalized cancer vaccine platform and are preparing to submit an FDA IND* with the goal of entering clinical trials in 2027.” He added, “We will enhance the scientific completeness of cancer vaccine development based on our proprietary AI technology and push forward the transition to the clinical stage step-by-step.”
*FDA IND: The procedure for obtaining permission from the U.S. Food and Drug Administration (FDA) to conduct clinical trials before administering a new drug to humans for the first time.
Dr. Jeong Yeon Kim and Dr. Jin Hyeon An participated as co-first authors in this study. The research results were published in the international scientific journal Science Advances on December 3rd.
※ Paper Title: B cell–reactive neoantigens boost antitumor immunity, DOI: 10.1126/sciadv.adx8303
2025.01.02 View 470