Retrieval-Augmented Long-Text Reasoning with Semantic Conflict Detection and Cross-Paragraph Evidence Integration

Abstract

This study addresses common challenges in long-text reasoning, including retrieval noise, dispersed evidence, semantic drift, and inconsistent generation, and proposes a retrieval-augmented generation framework for long-text understanding. The framework performs hierarchical semantic retrieval, cross-paragraph evidence graph modeling, semantic conflict detection and calibration, and evidence-guided generation to extract key segments from large-scale long texts, correct semantic relations, and produce reasoning outputs that align closely with factual content. In the retrieval stage, the model maps queries into a multi-scale semantic space and applies structured candidate filtering to improve cross-document evidence recall. In the evidence construction stage, graph structures represent semantic associations and potential conflicts across paragraphs, forming a unified semantic center that constrains subsequent generation. In the generation stage, evidence-aware attention and conditional control ensure that outputs rely strictly on real paragraph content, reducing hallucinations and enhancing the logical consistency of the reasoning chain. The experiments evaluate the model across answer accuracy, reasoning consistency, factual faithfulness, and cross-paragraph structural integration, and conduct sensitivity analyses on learning rate, context window length, retrieval candidate size, and training data scale. The results show that the framework achieves stable advantages across multiple metrics and generates high-quality reasoning outputs in complex long-text environments. Overall, this study provides a complete technical pathway that covers retrieval, integration, and generation, offering an effective solution for multi-source and multi-paragraph long-text tasks and establishing a methodological foundation for reliable long-text reasoning in knowledge-intensive application scenarios.