Large language models (LLMs) are powerful, but their massive size makes them computationally expensive. Researchers are constantly seeking ways to compress these models, making them more efficient without sacrificing too much performance. Traditional methods like quantization, pruning, and layer dropping often rely on simplifying assumptions, like the idea that compression errors simply add up across layers. However, this isn't always true, leading to suboptimal results, especially at high compression levels. A new research paper introduces EvoPress, a dynamic compression technique that uses an evolutionary search algorithm to find the optimal compression configuration for each part of the model. Imagine different parts of the model having different sensitivities to compression—some layers can be aggressively compressed with little impact, while others need to be treated more carefully. EvoPress automatically discovers this sensitivity by “evolving” better and better compression profiles. It starts with an initial compression setting and then generates slightly mutated versions. These “offspring” are evaluated by comparing their outputs to the original model, and the fittest survive to create the next generation. This iterative process quickly hones in on a highly optimized compression profile, outperforming previous methods, particularly at high compression ratios. EvoPress has been validated across various LLMs, including Llama, Mistral, and Phi models, and excels across different compression methods: layer dropping, unstructured sparsity, and quantization. It's even efficient enough to run on a single GPU, finding highly optimized configurations within hours. This evolutionary approach opens exciting doors for more efficient LLM deployment, making powerful AI more accessible and affordable. Future research directions include combining multiple compression methods into a single search and exploring finer-grained pruning techniques.