![Author ORCID: We display the ORCID iD icon alongside authors names on our website to acknowledge that the ORCiD has been authenticated when entered by the user. To view the users ORCiD record click the icon. [opens in a new tab]](https://www.cambridge.org/engage/assets/public/coe/logo/orcid.png){kind=logo}
Abstract
This paper presents the Pattern-Aware Complexity Framework (PACF), a novel approach that dynamically detects and leverages patterns during Large Language Model (LLM) generation to reduce computational complexity while maintaining output quality. Key contributions: - Achieves 93.8% average Pattern Utilization Efficiency (PUE) across diverse text types - Introduces two novel metrics: PUE and Pattern Harnessing Coefficient (PHK) for evaluating pattern-aware generation systems - Demonstrates intelligent content adaptation with PHK ranging from 13.7% (natural conversation) to 86.6% (code generation) - Maintains production-ready performance with <1% overhead and 10.7 tokens/second generation speed - Provides theoretical framework linking pattern detection to computational complexity reduction - Implements real-time pattern detection using n-grams, suffix trees, and attention patterns The framework was evaluated on 450 samples across six text categories (repetitive sequences, code, predictive patterns, random text, WikiText, and natural conversation) with statistical validation confirming significance (p < 10^-6 for speed, p < 10^-9 for perplexity). PACF is architecture-agnostic and compatible with existing transformer models, offering a 10× efficiency improvement that enables deployment on resource-constrained devices and reduces infrastructure costs for LLM applications. Open-source implementation available at: https://github.com/oliviersaidi/pacf-llm
Supplementary weblinks
Title
Pattern-Aware Complexity Framework (PACF): Official Software Implementation for Efficient LLM Generation
Description
This repository contains the complete open-source implementation of the Pattern-Aware Complexity Framework (PACF), as introduced in the accompanying publication. PACF is a novel, architecture-agnostic approach that dynamically detects and leverages recurring patterns during Large Language Model inference to achieve substantial computational reductions while preserving output quality. The software package includes modules for real-time pattern detection using n-gram models and suffix trees, integrated with transformer-based architectures. It enables reproduction of all key experimental results from the paper, including the demonstration of 93.8% Pattern Utilization Efficiency (PUE) and Pattern Harnessing Coefficient (PHK) measurements across six text categories (repetitive sequences, code, predictive patterns, random text, WikiText, and natural conversation). Key features include minimal runtime overhead (<1%), support for production deployment at 10.7 tokens/second, and compatibility with existing transformer models. The implementation provides researchers and practitioners with tools to benchmark pattern-aware generation techniques and deploy efficient LLM inference on resource-constrained devices.
Actions
View 
