Researchers have developed empirical scaling laws for compressing large language models (LLMs) in specific domains. They found that the performance of compressed LLMs degrades predictably as they are scaled down, but the rate of degradation depends on the type of supervision used during training. The study focused on quantitative finance and released a new dataset called FinHeadlineMix to aid in domain-specific compression decisions. --- Why it matters: These findings matter for researchers working with large language models because they provide a framework for making informed decisions about model compression, which is crucial for applications where latency and cost are critical. Source: https://arxiv.org/abs/2606.24747