KV-Cache Compression Methods Systematically Compared and Validated

This research undertakes a systematic comparison of two prominent KV-cache compression techniques: Turbo-Quant and SpectralQuant. The study employs a rigorous statistical validation methodology to differentiate between inherent codec differences and mere implementation variance. It also evaluates other non-dominated schemes, including WHT rotation with Beta Lloyd-Max and QJL, providing a comprehensive assessment of the landscape. A key finding is that eigenbasis-based compression methods, while powerful in structured data regimes, perform poorly when faced with heavy-tailed data due to issues with covariance instability. Conversely, these methods demonstrate strong performance where data is more structured. The research also highlights that the effective semantic dimension of the data adapts to the available calibration budgets rather than strictly reflecting the true data rank, offering a nuanced understanding of how these compression techniques operate under different constraints.