loading page

Factored Cognition Models: Enhancing LLM Performance through Modular Decomposition
  • Selma Wench,
  • Kyle Maxwell
Selma Wench

Corresponding Author:[email protected]

Author Profile
Kyle Maxwell
Author Profile

Abstract

Generative AI has achieved remarkable progress in producing human-like text, yet challenges remain in improving the predictability, reliability, and overall performance of language models. The novel approach of decomposing the GPT-2 model into specialized modules-memory, reasoning, and perception-significantly enhances its capabilities by compartmentalizing different cognitive functions. Comprehensive experiments demonstrate that the modular architecture results in substantial improvements in perplexity, accuracy, and F1 scores compared to baseline models. The findings demonstrate the potential of modular designs in achieving superior performance and adaptability, contributing to the field of factored cognition models. Addressing computational challenges and optimizing integration strategies will further enhance the practicality of deploying modular LLMs in diverse applications. The study provides a robust foundation for future research, exploring more efficient training algorithms, alternative modular configurations, and ethical considerations, ultimately advancing the development of sophisticated, reliable, and adaptable AI systems.
06 Jun 2024Submitted to TechRxiv
07 Jun 2024Published in TechRxiv