Sungkyunkwan University (SKKU) declared that an artificial intelligence research team from the College of Computing and Informatics, along with Professor Sung-Eun Hong and researchers Ae-cheon Jeong and Seung-hwan Lee, developed “SyMerge” by a joint study with NAVER AI Lab (Dr. Dong-yoon Han). The framework permit independently trained AI models to trade capabilities and increase overall performance while integrated into a single system.
The research findings were accepted and introduced at the 43rd International Conference on Machine Learning (ICML 2026).
Existing model-integrating strategies confronted technical bottlenecks in constructing multitasking AI. Merging models with different expertise often caused their knowledge to collide, outcomes in “task interference”—a phenomenon in which performance drops considerably compared with the original models. Until now, academia has centered closely on minimizing or preventing such conflict.
Shifting the paradigm, the research team targeted on creating synergy in which models actively supplement each other instead of simply avoid interference. The team identified that coordinating and optimizing the integrating ratio of one specific layer—the task-particular layer—among several internal layers can maximize compatibility among different models.
In specific, the newly evolved “SyMerge” technology announces an “Expert-Guided Self-Labeling” method. When encountering new, unlabeled data, the system trains itself by referencing the prediction capabilities of current models, which feature as experts. This permits the AI to navigate corrupted or altered data and maintain overall performance even under adverse conditions.
Furthermore, whilst conventional integrating strategies were limited to AI models derived from identical pretrained models, “SyMerge” successfully incorporates architectures with absolutely different pretrained origins—a feat previously deemed impossible.
Experimental results show that “SyMerge” obtains state-of-the-art (SOTA) performance throughout three core pillars of AI: image classification, computer vision-based prediction and natural language processing (NLP), showing its versatility and performance.
“This study represents a major milestone that shifts the paradigm of AI model integrating from ‘interference prevention’ to ‘mutual synergy creation,'” Hong stated. “By significantly decreasing the huge computing costs related with retraining AI, this technology will greatly make contribute to building lightweight yet highly versatile multitasking AI efficiently in the future.”
