Principles and effects of the AI model merging technology (SyMerge) developed by Professor Sung-Eun Hong's research team at Sungkyunkwan University. (Left) Process of AI knowledge merging and compatibility measurement, (Right) Performance effect achieved by adapting just a single core layer.
Sungkyunkwan University (SKKU) announced that an artificial intelligence research team from the College of Computing and Informatics (consisting of Professor Sung-Eun Hong and Researchers Ae-cheon Jeong and Seung-hwan Lee), through a joint study with NAVER AI Lab (Dr. Dong-yoon Han), has developed 'SyMerge.' This breakthrough framework allows independently trained AI models to trade capabilities and boost overall performance when merged into a single system.
Existing model merging techniques faced technical bottlenecks when attempting to build multi-tasking AI. Merging models with different expertise often caused their knowledge to collide, resulting in 'Task Interference'—a phenomenon where performance significantly drops compared to the original models. Until now, academia has focused heavily on merely minimizing or preventing such conflict.
Shifting the paradigm entirely, the SKKU research team focused on creating actual synergy where models actively complement each other instead of just avoiding interference. The team discovered for the first time that coordinating and optimizing the merging ratio of just one specific layer (the task-specific layer) out of numerous internal layers can maximize compatibility between different models.
In particular, the newly developed 'SyMerge' technology introduces an 'Expert-Guided Self-Labeling' method. When encountering new, unlabeled data, the system trains itself by referencing the prediction capabilities of the existing models, which function as experts. This enables the AI to navigate through corrupted or altered data and maintain highly stable performance even under adverse conditions.
Furthermore, while conventional merging techniques were restricted to AI models derived from identical pre-trained models, 'SyMerge' successfully integrates architectures with entirely different pre-trained origins—a feat previously deemed impossible.
Experimental results demonstrate that 'SyMerge' achieves State-of-the-Art (SOTA) performance across the three core pillars of AI: image classification, computer vision-based dense prediction, and natural language processing (NLP), proving both its versatility and superior performance.
"This study represents a major milestone that shifts the paradigm of AI model merging from 'interference prevention' to 'mutual synergy creation,'" said Professor Sung-Eun Hong of Sungkyunkwan University. "By drastically reducing the massive computing costs associated with retraining AI, this technology will greatly contribute to building lightweight yet highly versatile multitasking AI efficiently in the future."
The research findings have been accepted and presented at the 43rd International Conference on Machine Learning (ICML 2026), one of the world's most prestigious conferences in AI and machine learning, drawing significant interest from researchers globally.
