Phase transitions in large language model compression

Structural, numerical, and algebraic redundancy across pruning, quantization, and low-rank decomposition techniques are analyzed, enabling a criticality-aware compression framework that achieves near-lossless compression to 10% of the original size.