Explicit Semantic Weighting: How Prompt Repetition Modulates Output Depth in Large Language Models

Abstract

Traditional Prompt Engineering often prioritizes linguistic conciseness, structural elegance, and non-redundancy. However, empirical observations indicate that cleaner prompts do not systematically yield superior outputs. In this paper, we present an empirical investigation and theoretical framework exploring how Explicit Semantic Weighting (ESW)—achieved through explicit repetition and structural task decomposition within contexts—modulates the conditional probability distributions P(x | C) of Large Language Models (LLMs). Drawing analogies from human cognitive dynamics (e.g., the transition from Default Mode Network (DMN) to Executive Control Network (ECN) and local attention bias), we demonstrate that artificially introducing semantic density shifts an LLM’s vast, relatively uniform parametric knowledge retrieval toward localized, highly focused trajectories. Through two multi-turn comparative experiments (Presentation Generation and Open-Source Licensing Impact Analysis), we prove that: (1) task decomposition substantially constrains search space, and (2) explicit semantic repetition alters target attention weights, producing richer, deeper, and more human-satisfying responses despite lower linguistic prompt elegance. We formulate a novel benchmark matrix to quantify these cognitive shifts, paving the way for advanced LLM Task Orchestration.

Keywords