EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC 4.0

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

Felix Lin, Xinyi Lee, Nikhil Cohen, Lana Li

Computer Life · 28.10.2026 · 15 (1) · 30-36

DOI: 10.54097/vzmvf788 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Computer Life, online erschienen am 28.10.2026. Verfasst von Felix Lin und 3 weiteren Autor:innen. Die Arbeit nennt 28 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

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.

Zitieren

APA

Lin, F., Lee, X., Cohen, N., & Li, L. (2026). Query-Aware Cross-Segment KV Compression with Attention-Mass Calibration: An Operator-Level Study. Computer Life, 15(1), 30-36. https://doi.org/10.54097/vzmvf788

Vancouver

Lin F, Lee X, Cohen N, Li L. Query-Aware Cross-Segment KV Compression with Attention-Mass Calibration: An Operator-Level Study. Computer Life. 2026;15(1):30-36. doi:10.54097/vzmvf788

BibTeX
@article{Lin2026,
  author = {Lin, Felix and Lee, Xinyi and Cohen, Nikhil and Li, Lana},
  title = {Query-Aware Cross-Segment KV Compression with Attention-Mass Calibration: An Operator-Level Study},
  journal = {Computer Life},
  year = {2026},
  volume = {15},
  number = {1},
  pages = {30--36},
  doi = {10.54097/vzmvf788}
}
Kostenlos starten →