%43%68%61%6e%67%69%63%6b%20%4b%69%6d
-
KAIST Develops Key Technology to Make Personalized AI Safer
“Create an AI assistant trained only on our company’s documents.”
The era of building “personalized AI” by training AI models on individual or corporate documents and data is beginning. However, while such customization can improve task performance, it can also weaken the model’s existing safety safeguards. KAIST researchers have developed a core AI technology that preserves customized performance while further strengthening safety.
KAIST (President Choongsik Bae) announced on the 15th of July that a research team led by Professor Changick Kim from its School of Electrical Engineering has developed “Buffer-and-Reinforce,” a training framework for safe fine-tuning that prevents safety degradation when large language models (LLMs), such as
ChatGPT, are retrained on data from individuals or companies to better suit their needs.
Until now, one of the biggest challenges in the era of personalized AI has been that fine-tuning improves a model’s ability to perform new tasks, but can also weaken its existing safety rules. The research team focused on prior findings showing that, counterintuitively, fine-tuning an AI model while it is in a temporarily jailbroken state — a state in which it may respond even to dangerous requests it would normally refuse — does not significantly compromise its safety.
The team then devised a new approach in which this jailbroken state is not used in actual services, but is applied only temporarily during the fine-tuning process through a buffering module called “BufferLoRA,” which is removed after training.
The research team was the first to clarify why this phenomenon occurs. They found that, in the temporarily jailbroken state, the AI model becomes less easily influenced by harmful information, while still effectively learning the new task abilities desired by the user. In other words, the model can continue learning useful knowledge without additionally absorbing harmful behaviors.
Based on this insight, the team developed a two-stage learning method consisting of “buffering” and “safety reinforcement.”
First, the temporary buffering module, BufferLoRA, is applied to the AI model during user fine-tuning, where it acts as a protective layer that prevents harmful data from directly affecting the base model. Once fine-tuning is complete, this module is removed.
Next, a safety reinforcement module called “ReinforceLoRA” is applied to restore and strengthen the model’s safety. In this process, the team used QR decomposition, a mathematical technique that separates different types of information and selectively reflects only the necessary components. This allowed the model to retain the new functions learned from user data while selectively reinforcing safety.
Simply put, the researchers first placed a temporary protective layer, BufferLoRA, over the AI model so that harmful data could not directly affect it, while allowing the model to learn the necessary task. They then removed the protective layer and applied ReinforceLoRA to strengthen the model’s safety safeguards. As a result, the model maintained its customized performance while achieving even stronger safety.
In experiments, the AI model maintained high safety even in an extreme setting where all user data consisted of harmful questions and answers. After fine-tuning, the rate at which the AI generated harmful responses was about 8%, lower than the roughly 18% observed in the original model that had not been fine-tuned at all. The framework also achieved strong customized performance and state-of-the-art safety without requiring additional safety data during user fine-tuning or significantly increasing computational cost, suggesting its practical applicability to real-world personalized AI services.
Professor Changick Kim stated, “This research provides a key foundational technology that allows anyone to build customized AI with their own data while using it more safely,” adding, “We expect it to contribute significantly to building a trustworthy AI service environment in the era of personalized AI and AI agents.”
This research was led by Seokil Ham, a doctoral student in KAIST’s School of Electrical Engineering, as first author. The paper was selected as a Spotlight presentation at the International Conference on Machine Learning (ICML) 2026, one of the world’s most prestigious conferences in artificial intelligence, an honor given to only about the top 2.2% of all submitted papers, drawing international attention.
※ Paper title: Jailbreak to Protect: Buffering and Reinforcing via Temporary Jailbreaking for Safe Fine-Tuning in Large Language Models
DOI: 10.48550/arXiv.2605.24550
※ Author information: Seokil Ham (KAIST, first author), Jaehyuk Jang (KAIST, second author), Wonjun Lee (KAIST, third author), Changick Kim (KAIST, corresponding author)
※ Related video: https://drive.google.com/file/d/1gfok06dE8699qtiUR7gVsRoVmBGADaWQ/view?usp=sharing
This work was supported by Institute of Information & Communication Technology Planning & Evaluation (IITP) grant funded by Ministry of Science and ICT(MSIT) (No. RS-2025-02215344, Development of AI Technology with Robust and Flexible Resilience Against Risk Factors).
2026.07.15 View 1504 -
KAIST Illuminates the Eyes of Humanoid Robots with Minimal Memory
<CVPR 2026 poster session. From left to right: Minseok Seo (KAIST, first author), Mark Hamilton (MIT and Microsoft, second author), and Prof. Changick Kim (KAIST, corresponding author)>
From facial recognition on smartphones to humanoid robots, computer vision technology, which serves as the eyes of artificial intelligence (AI), is widely utilized in our daily lives. A joint research team from KAIST and international institutions has developed a technology that allows AI to see the world more clearly with minimal memory, increasing GPU (Graphics Processing Unit) memory efficiency by up to 16 times. This achievement is evaluated as a core technology that will accelerate the era of humanoid robots and on-device AI.
<Overview of Upsample Anything. Given a high-resolution image, it is first downsampled to a low-resolution image and then reconstructed through test-time optimization (TTO). During this process, pixel-wise anisotropic kernel parameters are learned. The learned kernels are subsequently applied to low-resolution foundation feature maps to generate high-resolution feature maps. These feature maps are then used to perform pixel-wise anisotropic Joint Bilateral Upsampling, enabling high-quality reconstruction at high resolution>
KAIST announced on June 17th that a research team led by Professor Changick Kim from the School of Electrical Engineering, through joint research with researchers from MIT and Microsoft in the United States, has developed 'Upsample Anything,' a universal technology that can enhance the visual performance of AI even with limited GPU memory.
Following its acceptance to 'CVPR 2026,' the world's most prestigious conference in the field of artificial intelligence and computer vision, this achievement was awarded the 'CVPR Compute Gold Star' in recognition of its efficient utilization of computational resources. It was also selected as the 'Transparency Champion,' ranking first overall in the category of research process transparency and reproducibility. This is an accomplishment that widely recognizes the core elements of responsible AI research, including research performance, computational resources used, code disclosure, and experimental reproducibility.
Recently, humanoid robots, autonomous driving systems, and AI based on world models (AI models that learn and predict the physical environment and changes of the real world) have been compressing input images into low-resolution features (core information extracted from images by AI) to increase computational speed and reduce memory usage.
However, during the compression process, a problem occurs where important visual information, such as small objects, thin structures, and minute defects, is lost. Conversely, processing all images at high resolution from the beginning requires massive GPU memory and computational resources, making real-time processing difficult. This has remained an unresolved challenge for a long time in situations where small devices like smartphones or robots, where mobility is crucial, must precisely perceive their surrounding environment.
To overcome these limitations, the research team developed a training-free (requiring no additional data training) upsampling technology that restores low-resolution feature information into high resolution by utilizing the edge and structural information of the input image.
Existing technologies required a separate retraining or complex optimization process to be applied to new environments or data. In contrast, 'Upsample Anything' developed by the research team can find the optimal restoration method using just a single input image, allowing it to be immediately applied to various environments.
In addition, by compressing and utilizing only core information instead of storing and processing all visual information at high resolution, GPU memory usage was significantly reduced. Based on a 224×224 size image (approximately 50,000 pixels) widely used in AI research, the research team restored visual information close to the original with a short calculation of about 0.4 seconds, achieving a performance that improves GPU memory efficiency by up to 16 times.
This means that artificial intelligence can perceive its surrounding environment more precisely even with limited computational resources. Therefore, this technology is expected to be widely used in various next-generation artificial intelligence fields, such as small devices like smartphones, as well as humanoid robots that need to accurately identify and manipulate small objects, autonomous driving systems, and on-device AI.
<Comparison image illustrating the performance gap with conventional methods (AI-generated). Conventional vision foundation models understand a scene by converting the input image into low-resolution features at a small patch level (left). Upsample Anything restores these low-resolution features to the original resolution level, enabling the AI to comprehend the scene's structure and boundaries with significantly higher precision (right)>
Professor Changick Kim said, “This technology is an algorithm that can significantly increase the visual precision of artificial intelligence with fewer resources, and it is expected to accelerate the commercialization of humanoid robots and on-device AI.” He added, “It is even more meaningful because it was recognized at CVPR not only for its performance but also for its computational efficiency and research transparency.”
This research was participated in by KAIST PhD student Minseok Seo as the first author, and this achievement was presented on June 7 at 'CVPR 2026,' the world's most prestigious conference in the field of artificial intelligence and computer vision.
※ Paper Title: Upsample Anything: A Simple and Hard to Beat Baseline for Feature Upsampling, DOI:10.48550/arXiv.2511.16301
※ Author Information: Minseok Seo (KAIST, First Author), Mark Hamilton (MIT, Microsoft, Second Author), Changick Kim (KAIST, Corresponding Author)
2026.06.17 View 2963