Machine learning researchers solve a 15-year-old problem in continual learning with a tiny add-on

Researchers found a way to let AI models keep learning from new data without forgetting old lessons, using lightweight plugin components instead of redesigning the entire system. The fix adds only 4% more parameters to a model but improves accuracy by 8%, which means the same computational cost buys you better performance.

Continual learning is a hard problem in deployed AI systems — models need to adapt to new data without erasing what they already learned, and nobody has had a good solution that doesn't require massive computational overhead. This paper shows that the bottleneck wasn't the overall architecture, it was the coupling between learning new things and preserving old ones. If this approach scales, it means AI systems that learn continuously in production become cheaper to run and less likely to catastrophically forget what matters.