Model Compression

Model compression studies how to make trained models smaller or cheaper to run, especially for large language models, vision and generative models, and deployment on edge devices. Common approaches include quantization, pruning, low-rank factorization, layer or channel merging, and knowledge distillation; some methods combine them or optimize compression for a specific task and hardware. Evaluation typically measures model size, inference speed, and task quality on benchmarks such as language-model perplexity, vision accuracy, and generation quality. A central challenge is that parameter reduction does not always yield faster deployment or preserve reasoning, fairness, safety, and adaptability, so hardware-aware and task-specific evaluation remains important.