Scientific papers are lengthy, with research points scattered across sections. Existing automatic summarization methods rely solely on sentence weights and paragraph positions for text selection, which easily leads to the loss of key information and causes logical breaks in summaries. While general large language models generate fluent summaries, they suffer from difficulties in constraining output structure and factual inaccuracies. To address these issues, this study proposes a move-guided structured summarization model for scientific papers—Move-driven Summarization (MoverSum). MoverSum employs a Seq2Seq multi-task learning framework, using fine-grained moves as global planning cues. It integrates text semantics, section positions, and move labels to construct multi-dimensional features, simultaneously accomplishing sentence importance evaluation and summary move sequence prediction. Key sentences are then extracted according to the move sequence to generate a well-organized structured summary. Comprehensive experiments based on arXiv and PubMed datasets show that MoverSum outperforms various baseline models across all ROUGE metrics. Human evaluation further confirms that summaries generated by this model have more complete content coverage and smoother writing logic. Ablation experiments and comparisons with large models demonstrate that multi-dimensional features and move planning are core to the model's performance improvement, and MoverSum is better adapted to scientific text summarization scenarios. The proposed method balances the factual fidelity of extractive summaries with the controllability of output structure, offering a novel move-aware modeling approach for the task of structured summarization of scientific papers.
