A new framework called CHEEM allows AI models to learn new tasks without sacrificing performance on previously learned tasks. This framework improves operational efficiency by reducing the number of computational steps required for simple tasks.
“CHEEM addresses two long-standing challenges for AI models: continuous learning and adaptive intelligence,” said Tianfu Wu, associate professor of computer engineering at North Carolina State University. Continuous learning allows the model to incorporate new data and tasks, but it often degrades performance on previous tasks.
Adaptive intelligence involves changing the computational process based on the complexity of the task. Many AI models utilize the same computational chain regardless of the task, which can be inefficient. “We believe that these two challenges are intertwined, and that improving the continuous learning capabilities of models can advance toward adaptive intelligence. This is the fundamental idea behind CHEEM,” Wu said.
CHEEM stands for Continual Hierarchical-Exploration-Exploitation Memory and provides the flexibility to leverage existing computational architectures, allowing models to change, skip, or add layers when faced with new tasks. This design helps maintain existing knowledge while integrating new data and managing computational resources depending on the complexity of the task.
To evaluate CHEEM, researchers used state-of-the-art vision transformer models on two challenging benchmark datasets: MTIL and VDD. Wu said benchmarks are “good test cases” because of their complexity and variety.
CHEEM significantly outperforms existing continuous learning methods across both benchmarks. Wu said, “CHEEM came very close to achieving a full fine-tuning upper bound on these new tasks, meaning it was about as good as if we had trained the model to perform just that one task.”
The framework also improved the adaptive intelligence of the model, adjusting its computational structure to the complexity of the task. This model has semantically adjusted its architecture, adopting existing layers for tasks similar to the previous one and adding new layers for tasks that are clearly different. “We are excited about what we were able to demonstrate at CHEEM,” Wu said.
The researchers are currently seeking collaborators to access the computational resources needed to evaluate CHEEM on large-scale underlying models with billions of parameters. A peer-reviewed paper detailing CHEEM will be presented at the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2026), June 3-7 in Denver, Colorado.
Featured image credit
