CAS trains composable per-document KV cache cartridges via dynamic distractor mixing and a rotating budget manager, scaling to million-token collections with 10-31 point gains over monolithic cartridges and matching RAG at 3-4x lower token cost.
Maxime Louis, Hervé Déjean, and Stéphane Clin- chant
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Cartridges at Scale: Training Modular KV Caches over Large Document Collections
CAS trains composable per-document KV cache cartridges via dynamic distractor mixing and a rotating budget manager, scaling to million-token collections with 10-31 point gains over monolithic cartridges and matching RAG at 3-4x lower token cost.