AuthorityKV: Preventing Authority Reversals in KV-Cache Compression
Proceedings of EACL 2027 (main conference), to appear
Shows that KV-cache compression can silently break a prompt's instruction hierarchy: across three instruction-tuned models, three widely used compressors reverse 2,055 of 7,200 decisions that the uncompressed models get right. AuthorityKV, a training-free rule for any score-based compressor at an unchanged budget, removes all of these reversals with annotated span boundaries and 91% of them using only the chat roles and source provenance an orchestrator already holds; it raises system-prompt compliance on System-IFEval by 36–46 points and cuts successful prompt injections from 7 of 64 to 0.
@inproceedings{zhang2027authoritykv,
title = {{AuthorityKV}: Preventing Authority Reversals in
{KV}-Cache Compression},
author = {Zhang, Xingjian},
booktitle = {Proceedings of the European Chapter of the Association
for Computational Linguistics (EACL 2027)},
year = {2027},
publisher = {Association for Computational Linguistics},
note = {To appear}
}