Rethinking Compression: Reduced Order Modelling of Latent Features in Large Language Models