Query-Aware Cross-Segment KV Compression with Attention-Mass Calibration: An Operator-Level Study

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Cross-segment KV compression reduces the history accessible during masked denoising, but the number of summaries retained from each completed segment can distort attention mass. We implement the learned-query CSKV pooling operation of SW-SpeedDLM using Q/K/V extracted from an actual four-head bidirectional masked encoder, then examine query-aware integer budget allocation and an analytic log(W/c) multiplicity correction. Evaluation uses 170 held-out 320-character spans from an Opticks transcription and a Wikipedia extract. With four 64-character segments, uniform pooling at the compressor’s training budget (B=16) is hard to beat with prefix-based allocation (NMSE 0.0054 versus 0.0050), whereas at untrained budgets such as B=8 adaptive allocation avoids the collapse of uniform pooling (0.0389 versus 1.8093). With query-ranked slots, calibrated adaptive routing at B=16 achieves 0.0044 versus 0.0051 without calibration; however, a paired interval for its difference against uniform pooling crosses zero. Further independently initialized and variable-budget-trained compressors show that allocation effects depend on learned bank quality. The block-constant attention identity is verified numerically and final-bank storage is exactly 93.75% smaller at B=16 than 256 uncompressed positions. The experiment measures cross-window attention fidelity, not end-to-end diffusion generation or GPU throughput.