Instead of training a massive model and shrinking it afterward, CompreSSM compresses the model while it is still learning

https://techxplore.com/news/2026-04-compression-technique-ai-leaner-faster.html

"shrinking AI expensive/ time-consuming: removing parts after training/ training small to copy big... overcome using state-space models identifying parts of model pulling their weight early in training/ removing dead weight... reverts to a previous checkpoint if performance drops... compressed models train up to 1.5X faster than full-sized... reduced to 25% original size 85.7% accurate, outperforming same small model trained from scratch... 40X faster and more accurate than other modern spectral techniques"