Cache-Aware Generative NLP: Reducing Latency in Large-Scale Multilingual Systems.
Students & Supervisors
Student Authors
Supervisors
Abstract
Large-scale multilingual generative models face significant computational bottlenecks due to repetitive linguistic computations across languages. While existing caching mechanisms focus on token-level optimization, they overlook the computational redundancy in morphological and syntactic processing that occurs during multilingual text generation. We introduce Adaptive Linguistic Computation Caching (ALCC), a novel framework that identifies and caches expensive linguistic transformation computations during the generation process. ALCC employs a three-tier caching architecture: (1) Morphological Transformation Cache for expensive inflection computations, (2)Syntactic Pattern Cache for grammatical structure generation,and (3) Cross-Language Transfer Cache that leverages typological similarities to share computational patterns between related languages. Our approach introduces a Linguistic Complexity Predictor that dynamically identifies high-cost linguistic operations and prioritizes them for caching. Experiments on morphologically rich languages (Finnish, Turkish, Arabic) show an average of 75.8% latency reduction across 12 models on curated settings; gains vary with typological similarity and warm-up. in generation tasks while maintaining identical output quality. This work represents the first systematic approach to caching linguistic computations in generative multilingual systems.
Keywords
Publication Details
- Type of Publication:
- Conference Name: 2026 IEEE 2nd International Conference on Quantum Photonics, Artificial Intelligence & Networking (QPAIN 2026)
- Date of Conference: 16/04/2026 - 16/04/2026
- Venue: IT Business Incubator, Chittagong University of Engineering and Technology (CUET), Chattogram, Bangladesh
- Organizer: IEEE Photonics Society Bangladesh Chapter