Title: MemoNoveltyAgent: A Workflow for Assessing Paper Novelty with Historical Research Memory

Abstract:

To mitigate the overwhelming workload associated with screening academic papers, scholars are increasingly turning to general-purpose AI agents, including AI reviewers and tools like DeepResearch, for evaluating novelty. However, because these systems lack specialized mechanisms for handling scholarly literature, their outputs often suffer from superficiality and noticeable quality gaps. To address this limitation, we present MemoNoveltyAgent, a multi-agent framework designed to produce comprehensive and accurate novelty reports. In addition to retrieving specific prior-paper evidence through Retrieval-Augmented Generation (RAG), our approach integrates a high-level abstract memory derived from large-scale scholarly corpora. This memory structures research into hierarchical trees to extract field-specific evolutionary trajectories, offering a broader historical perspective. The system further breaks down papers into distinct novelty points to facilitate fine-grained analysis and retrieval, utilizing a self-validation mechanism to enhance the faithfulness of the generated reports. To overcome the evaluation difficulties inherent in open-ended generation tasks, we introduce a RAG-augmented checklist evaluation method that ensures reliable, evidence-based assessments. Our extensive experiments show that MemoNoveltyAgent surpasses GPT-5 DeepResearch by 13.69%. The code and demo are accessible at https://github.com/SStan1/MemoNoveltyAgent

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC