VestigeKV holds 1.00 retrieval at 32x compression by evicting on a vestigial RoPE branch.

Attention-observed selectors like H2O and SnapKV collapse to 0.00-0.33 needle retrieval on a NoPE MLA model, because a long-lived cache must be compressed before the queries that will read it exist. On Kimi Linear, VestigeKV evicts by a query-independent signal already in the cache, the 64-dimensional decoupled branch that NoPE training repurposes from a RoPE vestige into a salience channel, reading 11% of each row and moving non-top rows into a GPU-resident archive. No training, no quantization, no kernel change. Retrieval holds 1.00 at 8x and 0.92 at 32x from 8k to 65k. The identical operator on a RoPE MLA collapses to 0.08, so the effect is NoPE-exclusive. arXiv 2609.03949