MVSS: A Unified Framework for Multi-View Structured Survey Generation

ACL ARR 2026 May Submission14718 Authors

Keywords: automatic survey generation, structured text generation, large language models, hierarchical knowledge structures

Abstract: Scientific surveys demand not just summarization but the organization of literature into coherent conceptual structures. Existing automatic survey generation methods, however, focus on linear text generation and largely ignore hierarchical relations and structured methodological comparisons, leaving substantial gaps to expert-written surveys. Evaluation has also been confined to ad-hoc CS topic lists, leaving cross-domain generalization untested. We address these limitations along two complementary lines. First, we propose MVSS, a constraint-driven multi-view survey generation framework: it constructs a citation-grounded Hierarchical Knowledge Tree (HKT) that serves as a strict anchor for tree-induced comparison tables, outline planning, and narrative writing, followed by a cross-view alignment step that resolves coverage, table-narrative binding, citation, and traversal inconsistencies. Second, we construct MVS-Bench, the first cross-domain benchmark for automated survey generation, comprising 100 expert-curated topics across computer science, economics, electrical engineering, and systems science, paired with high-impact reference surveys and a multi-signal evaluation suite. On MVS-Bench, MVSS significantly outperforms existing methods in survey organization and evidence grounding across all four disciplines, and substantially narrows the gap to expert-written surveys. We open our resources at https://anonymous.4open.science/r/MVSS-824F.

Paper Type: Long

Research Area: LLM agents

Research Area Keywords: Generation, Resources and Evaluation, Machine Learning for NLP

Contribution Types: NLP engineering experiment

Languages Studied: English

EMNLP 2026 AI Reviewing Experiment: yes

Submission Number: 14718