Structure-preserving Pruning with Recoverable Capacity for Efficient Large Language Model Serving
International Journal of Business and Social Science Studies · 22.09.2026 · 6 (2) · 248-260
DOI: 10.66795/ijbsss76 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Business and Social Science Studies, online erschienen am 22.09.2026. Verfasst von Julia Taylor und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The rapid proliferation of large language models has significantly advanced the field of natural language processing but has concurrently introduced unprecedented challenges regarding computational resource requirements and serving latency. Traditional model compression techniques, such as unstructured pruning, often fail to translate theoretical sparsity into tangible latency reductions due to the irregular memory access patterns they create. Conversely, structured pruning alters the fundamental architecture of the model, complicating deployment and often resulting in irreversible capacity degradation that severely impacts complex reasoning tasks. This paper introduces a novel paradigm termed structure-preserving pruning with recoverable capacity, designed specifically to optimize the serving efficiency of large language models while maintaining their inherent architectural integrity. By logically masking redundant pathways rather than physically removing them, the proposed methodology preserves the dense tensor structure required for optimal hardware utilization. Furthermore, we introduce a dynamic routing mechanism that evaluates token complexity in real-time, allowing the model to recover dormant capacity for demanding inputs while processing simpler inputs through the accelerated pruned pathways. Comprehensive empirical evaluations demonstrate that this approach achieves significant improvements in serving throughput and generation latency without the catastrophic quality degradation typically associated with static pruning methodologies.
Zitieren
APA
Taylor, J., & Collins, J. J. (2026). Structure-preserving Pruning with Recoverable Capacity for Efficient Large Language Model Serving. International Journal of Business and Social Science Studies, 6(2), 248-260. https://doi.org/10.66795/ijbsss76
Vancouver
Taylor J, Collins JJ. Structure-preserving Pruning with Recoverable Capacity for Efficient Large Language Model Serving. International Journal of Business and Social Science Studies. 2026;6(2):248-260. doi:10.66795/ijbsss76
BibTeX
@article{Taylor2026,
author = {Taylor, Julia and Collins, Jack J.},
title = {Structure-preserving Pruning with Recoverable Capacity for Efficient Large Language Model Serving},
journal = {International Journal of Business and Social Science Studies},
year = {2026},
volume = {6},
number = {2},
pages = {248--260},
doi = {10.66795/ijbsss76}
}
EUVIMED