Transactional Attention: Semantic Sponsorship for KV-Cache Retention

Abhinaba Basu · 2026

Read Transactional Attention: Semantic Sponsorship for KV-Cache Retention by Abhinaba Basu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-11288.txt.

Description

Basu tackles a blind spot in KV-cache compression: dormant tokens such as credentials and keys draw almost no attention yet prove essential at generation time. At just 16 tokens, a mere 0.4 percent of a 4K context, every existing method scores zero on credential retrieval. His Transactional Attention lets structural anchors like 'key:' shield nearby values from eviction, reaching 100 percent retrieval where six baselines score zero, and a fast variant cuts memory overhead by 52 percent.

Catalog record

CREATORAbhinaba Basu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38413
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.11288

Browse the full library · About the Tower of Babel