Structured Path Guidance for Logical Coherence in Large Language Model Generation

Abstract

 In this paper, a thought chain control generation method based on structure guidance is proposed to solve the problem that large language models lack logical control and structural constraints in complex language generation tasks. This method introduces a structure encoding module and a dynamic structure state mechanism to guide the model to develop a thought chain along a preset structural path during the generation process, thereby improving the logical coherence and structural consistency of the generated content. Specifically, the model first converts the task requirements into a structured representation, fuses it with the input semantic context to construct a joint representation, and then generates it step by step through a structure-aware decoder. In this process, the dynamic structure state updates and constrains the generation state to achieve real-time regulation of the language output path. In order to systematically evaluate the effectiveness of this method, this paper designs experiments in multiple dimensions, including structural step sensitivity, comparison of structural representation methods, and performance comparison with a variety of public models. A comprehensive test is carried out using the NarrativeQA dataset. The experimental results show that this method is significantly superior to existing mainstream methods in terms of structural alignment, thought chain coherence, and generation accuracy, which effectively verifies the control value and modeling advantages of the structure guidance mechanism in generation tasks.