A new framework called CHEEM enables AI models to learn new tasks without sacrificing performance on previously learned tasks. The framework enhances operational efficiency by reducing the number of computational steps required for simpler tasks.
Tianfu Wu, an associate professor of computer engineering at North Carolina State University, said, “CHEEM addresses two longstanding challenges for AI models: continual learning and adaptive intelligence.” Continual learning allows models to incorporate new data and tasks, but often leads to degradation in performance on earlier tasks.
Adaptive intelligence involves modifying computation processes based on task complexity. Many AI models utilize the same computation chain regardless of task, which can be inefficient. Wu stated, “We think these two challenges are intertwined, and that we can make progress toward adaptive intelligence by improving a model’s ability to engage in continual learning. This is the fundamental idea behind CHEEM.”
CHEEM, short for Continual Hierarchical-Exploration-Exploitation Memory, provides flexibility in utilizing existing computational architecture, enabling models to modify, skip, or add layers when faced with new tasks. This design helps maintain existing knowledge while integrating new data and managing computational resources according to task complexity.
To evaluate CHEEM, the researchers used a state-of-the-art vision transformer model on two challenging benchmark datasets: MTIL and VDD. Wu described the benchmarks as “good test cases” due to their complexity and variety.
CHEEM significantly outperformed existing continual learning methods across both benchmarks. Wu noted, “CHEEM got very close to achieving the full fine-tuning upper bound for these new tasks, meaning that it was almost as good as if you had trained the model to only perform that one task.”
The framework also improved the model’s adaptive intelligence, aligning its computational structure with task complexity. The model tailored its architecture semantically, employing existing layers for tasks similar to prior ones and adding new layers for distinctly different tasks. “We’re excited about what we’ve been able to demonstrate with CHEEM,” Wu said.
The researchers are currently seeking collaborators to access computational resources necessary for evaluating CHEEM on large foundation models with billions of parameters. The peer-reviewed paper detailing CHEEM will be presented at the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2026) from June 3–7 in Denver, Colorado.
