Automating structured reasoning with entity-driven semantic simplification in Llms

Chain-of-Thought (CoT) prompting has substantially improved zero-shot reasoning in large language models (LLMs). However, for problems involving multiple interacting entities, entangled dependencies and scattered attribute information often hinder semantic understanding, which in turn degrades downstream reasoning accuracy. To address this issue, we propose Entity-Driven Semantic Simplification Prompting (EDSSPrompt), a zero-shot prompting framework that introduces a semantic preprocessing stage before reasoning. Specifically, EDSSPrompt extracts key entities and their attributes from the original problem, restructures them into a linearized representation, and reformulates the input to better align with CoT’s step-by-step reasoning pattern. Experiments across three open-source LLMs show that EDSSPrompt consistently improves reasoning accuracy, achieving average gains of [Formula: see text], [Formula: see text], and [Formula: see text] over standard zero-shot CoT prompting. These results demonstrate that EDSSPrompt provides a lightweight, modular, and training-free solution for enhancing reasoning performance, making it well suited for practical deployment in resource-constrained AI systems. Our code will be made publicly available at https://anonymous.4open.science/r/EDSS-25A5/README.md.