Cognitive Chain-Based Dual Fusion Framework for Multi-Document Summarization

Multi-Document Summarization (MDS) is a critical task in natural language processing that aims to condense document clusters into concise and comprehensive summaries. However, existing approaches based on large language models (LLMs) often lack structured quality monitoring and depth refinement mechanisms. This opacity and lack of self-correction can compromise the reliability, depth, and controllability of the resulting summaries. To address these limitations, this paper introduces the Self-Optimizing Multi-Path Fusion Framework (MFOG), a novel conceptual architecture for MDS. MFOG treats MDS as a collaborative process that optimizes both summary depth and breadth. The framework uses a dual-path architecture to balance summary depth and breadth. A depth-focused path, augmented by Retrieval-Augmented Generation (RAG), progressively enhances content depth and logical coherence. Concurrently, a breadth-first parallel path ensures comprehensive coverage. A final fusion module then performs a weighted integration of these outputs. We present an illustrative experimental study on benchmark datasets. On Multi-News, MFOG achieves ROUGE-1 and ROUGE-2 scores of 51.08 and 22.76, representing improvements of 1.23 and 1.11, respectively, over the strongest baselines.On DUC-2004, it achieves a ROUGE-1 score of 36.12 (a 1.30 improvement) and a BERTScore of 40.16 (a 1.14 improvement). This preliminary study validates the feasibility of the MFOG framework, demonstrating its potential to produce summaries that are both comprehensive and coherent.