Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Yinqi Liu, Yueqi Zhu, Yongkang Zhang, Xinfeng Li, Feiran Liu, Yufei Sun, Xin Wang, Renzhao Liang, Yidong Wang, Cunxiang Wang

TL;DR

MVSS introduces a structure-first framework for multi-view structured survey generation that jointly builds a citation-grounded Hierarchical Knowledge Tree (HKT), tree-induced structured tables, and narrative text guided by cross-view alignment. It comprises three stages—HKT construction, table generation conditioned on the tree, and outline/text generation constrained by both structures with an alignment step to resolve conflicts. Evaluated on 76 computer science topics with a 530k arXiv corpus, MVSS outperforms automatic baselines in coverage, structure, and relevance, achieving near-expert quality at longer lengths and reliable citation grounding (Rec/Prec > 75%). The results demonstrate that explicit hierarchical organization and cross-view constraints can substantially improve the quality and grounding of automated surveys, suggesting a scalable path toward expert-like, structured literature synthesis.

Abstract

Scientific surveys require not only summarizing large bodies of literature, but also organizing them into clear and coherent conceptual structures. Existing automatic survey generation methods typically focus on linear text generation and struggle to explicitly model hierarchical relations among research topics and structured methodological comparisons, resulting in gaps in structural organization compared to expert-written surveys. We propose MVSS, a multi-view structured survey generation framework that jointly generates and aligns citation-grounded hierarchical trees, structured comparison tables, and survey text. MVSS follows a structure-first paradigm: it first constructs a conceptual tree of the research domain, then generates comparison tables constrained by the tree, and finally uses both as structural constraints for text generation. This enables complementary multi-view representations across structure, comparison, and narrative. We introduce an evaluation framework assessing structural quality, comparative completeness, and citation fidelity. Experiments on 76 computer science topics show MVSS outperforms existing methods in organization and evidence grounding, achieving performance comparable to expert surveys.

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

TL;DR

MVSS introduces a structure-first framework for multi-view structured survey generation that jointly builds a citation-grounded Hierarchical Knowledge Tree (HKT), tree-induced structured tables, and narrative text guided by cross-view alignment. It comprises three stages—HKT construction, table generation conditioned on the tree, and outline/text generation constrained by both structures with an alignment step to resolve conflicts. Evaluated on 76 computer science topics with a 530k arXiv corpus, MVSS outperforms automatic baselines in coverage, structure, and relevance, achieving near-expert quality at longer lengths and reliable citation grounding (Rec/Prec > 75%). The results demonstrate that explicit hierarchical organization and cross-view constraints can substantially improve the quality and grounding of automated surveys, suggesting a scalable path toward expert-like, structured literature synthesis.

Abstract

Scientific surveys require not only summarizing large bodies of literature, but also organizing them into clear and coherent conceptual structures. Existing automatic survey generation methods typically focus on linear text generation and struggle to explicitly model hierarchical relations among research topics and structured methodological comparisons, resulting in gaps in structural organization compared to expert-written surveys. We propose MVSS, a multi-view structured survey generation framework that jointly generates and aligns citation-grounded hierarchical trees, structured comparison tables, and survey text. MVSS follows a structure-first paradigm: it first constructs a conceptual tree of the research domain, then generates comparison tables constrained by the tree, and finally uses both as structural constraints for text generation. This enables complementary multi-view representations across structure, comparison, and narrative. We introduce an evaluation framework assessing structural quality, comparative completeness, and citation fidelity. Experiments on 76 computer science topics show MVSS outperforms existing methods in organization and evidence grounding, achieving performance comparable to expert surveys.
Paper Structure (26 sections, 12 equations, 4 figures, 11 tables)

This paper contains 26 sections, 12 equations, 4 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Automatic Survey Generation
  4. Knowledge Structuring and Multi-Signal Evaluation
  5. Method
  6. Evidence-Grounded Tree (HKT)
  7. Tree-Induced Table Generation
  8. Tree+Table Guided Outline and Survey
  9. Experiments
  10. Experimental Setup
  11. Dataset and Corpus.
  12. LLM-as-judge evaluation.
  13. Metrics
  14. Survey quality.
  15. Tree quality.
  16. ...and 11 more sections

Figures (4)

  • Figure 1: Motivation for structure-first survey generation. Linear-text survey generation obscures evolving comparison dimensions, motivating aligned multi-view structures. Expert surveys predominantly adopt explicit hierarchies (2021--2025), supporting a structure-first perspective.
  • Figure 2: Overview of MVSS. Given a topic $T$ and a paper database $D$, MVSS constructs an evidence-grounded hierarchical tree, generates aligned comparison tables, and produces a structured survey via cross-view alignment.
  • Figure 3: Double-blind pairwise human evaluation results. The figure shows win/tie/loss counts for MVSS compared with AutoSurvey and human-written surveys.
  • Figure 4: Performance evolution across iterative tree refinement rounds. Iter 0 corresponds to the initial tree without refinement.