Sparse KV compression is a common way to cut the cost of long-context inference: instead of keeping every past state, the model retains only a selected subset. But those selected states usually keep the positional coordinates they had in the full sequence. That choice, the paper argues, is a problem when the model uses Rotary Position Embedding (RoPE).

RoPE encodes position as a rotation, so the meaning of a state depends on where it sat relative to its neighbours. When compression removes tokens, the remaining states are left with gaps in their positional coordinates. Those gaps distort the geometric relationships that RoPE relies on, even though the selected states themselves are unchanged.

The paper's proposal is to recompute positional geometry after selection rather than before it. By assigning new coordinates to the retained states based on their post-compression arrangement, the model can avoid the distortion introduced by the gaps. The abstract does not provide experimental results, so the effectiveness of the approach is not detailed in the source.