Online adaptation of KV-cache compression to non-stationary marginal utility
Develop online updates to the compression basis of the Key–Value cache so that the cache budget can be adaptively reallocated as new tokens are processed and the eigenvalue spectrum of the projected covariance operator changes during generation, particularly for borehole logs whose information-dense tokens are distributed throughout the sequence.
References
Four research directions, identified in the original formulation, remain open and are sharpened by the application context.
\paragraph{Non-stationary marginal utility.} The eigenvalue spectrum of $W V W{\top}$ changes during a generation sequence as new tokens are processed. Static compression schemes based on the initial spectrum will be suboptimal for long-context generation. In the geo-mining context, this is particularly acute for borehole logs, where the most information-dense tokens (depth measurements, mineralogical observations) are distributed throughout the sequence rather than concentrated at the beginning. Online updates to the compression basis, analogous to online PCA (principal component analysis performed incrementally on a data stream: the eigenbasis is updated with each new observation via a rank-one correction, avoiding the need to store or reprocess past data \citep{oja1982}), could adaptively reallocate the KV cache budget as new tokens are processed.