cs.CLMar 28, 2026

Story2Proposal: A Scaffold for Structured Scientific Paper Writing

Authors: Zhuoyang Qian, Wei Shi, Xu Lin, Li Ling, Meng Luo, Ziming Wang, Zhiwei Zhang, Tengyue Xu, 

Organizations: AgentAlpha Team.

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Generating scientific manuscripts requires maintaining alignment between narrative reasoning, experimental evidence, and visual artifacts across the document lifecycle. Existing language-model generation pipelines rely on unconstrained text synthesis with validation applied only after generation, often producing structural drift, missing figures or tables, and cross-section inconsistencies. We introduce Story2Proposal, a contract-governed multi-agent framework that converts a research story into a structured manuscript through coordinated agents operating under a persistent shared visual contract. The system organizes architect, writer, refiner, and renderer agents around a contract state that tracks section structure and registered visual elements, while evaluation agents supply feedback in a generate evaluate adapt loop that updates the contract during generation. Experiments on tasks derived from the Jericho research corpus show that Story2Proposal achieved an expert evaluation score of 6.145 versus 3.963 for DirectChat (+2.182) across GPT, Claude, Gemini, and Qwen backbones. Compared with the structured generation baseline Fars, Story2Proposal obtained an average score of 5.705 versus 5.197, indicating improved structural consistency and visual alignment.

Figures & tables

Figure 1: System overview of Story2Proposal. Given a research story as input, the framework coordinates architect, writer, refiner, and renderer agents through a persistent shared visual contract that records section structure, registered visual artifacts, and validation rules. Evaluation agents inspect intermediate outputs and feed corrective signals back to the contract state, enabling provenance-aware planning, structurally constrained drafting, global refinement, and deterministic rendering of the final manuscript.
Figure 2: Schema of the shared visual contract used by Story2Proposal. The contract combines a global registry of figures, tables, and citation slots with section-level obligations and validation rules, so that all agents operate on the same structural state and the renderer can deterministically verify label uniqueness, reference resolution, and narrative–visual alignment before compilation.
Figure 3: Expert evaluation protocol used in our experiments. Each condition combines one of two generation methods, one of four LLM backbones (GPT, Claude, Gemini, and Qwen), and ten independent expert reviewers. Reviewers assess complete manuscripts on structural integrity, writing clarity, methodological rigor, experimental substance, citation hygiene, reproducibility, formatting stability, and visual communication, and the resulting scores are aggregated into the overall expert evaluation metric.
ModelE1E2E3E4E5E6E7E8E9E10Avg
Based on DirectChat
GPT3.623.015.973.333.494.065.344.233.644.094.08
Claude3.392.685.983.043.363.964.883.923.234.203.86
Gemini3.322.796.073.163.513.665.143.993.564.143.93
Qwen3.382.746.193.223.413.935.324.053.583.933.98
Based on Story2Proposal
Table 2: Average expert evaluation scores comparing Story2Proposal and DirectChat across four LLM settings. Higher is better.
PaperE1E2E3E4E5E6E7E8E9E10Avg
Based on Fars
Escrowed Batch Reveal4.554.396.064.414.385.046.194.574.475.224.93
Symbolic Execution4.975.646.325.405.655.576.465.115.575.695.64
Hazard-Signature Tombstones4.274.066.094.194.334.576.044.594.214.744.71
Poisoning LLM-Induced Rules5.005.596.015.115.355.226.805.265.145.635.41
Based on Story2Proposal
Table 3: Average expert evaluation scores comparing Story2Proposal and Fars across four benchmark research papers. Higher is better.
Figure 4: Average performance deltas between Story2Proposal and baseline generation methods across experiments. Positive values indicate that the contract-governed multi-agent framework outperforms the comparison system under the same backbone or benchmark setting. The figure summarizes both the large gains over DirectChat in the cross-model study and the smaller but consistent improvements over the structured Fars baseline, highlighting the benefit of stronger structural control and narrative–visual alignment.
More in this topic

Long-Form Document Generation

Long-form document generation studies methods for producing coherent, factual, well-structured books, reports, and other extended texts with AI.

See all 32 papers

Explore similar work

CardsList

  1. PaperOrchestra: A Multi-Agent Framework for Automated AI Research Paper Writing

    Apr 6, 2026Yiwen Song, Yale Song, Tomas Pfister +1AI-Assisted Scientific ResearchAI Agent Benchmarks

    Google

  2. LECTOR: Joint Optimization of Scientific Reasoning Graphs and Introduction Generation

    May 25, 2026Jiabei Xiao, Yizhou Wang, Chen Tang +3RL for Language Model ReasoningAI-Assisted Scientific Research

    1The Chinese University of Hong Kong · 2Shanghai Artificial Intelligence Laboratory · 3Fudan University

  3. Prompt-to-Paper: Agentic AI System for Bioinformatics

    Jul 5, 2026Ramsha Kamran, Maheera Amjad, Zartasha Mustansar +3Long-Form Document GenerationAI-Assisted Scientific Research

    School of Interdisciplinary Engineering and Sciences (SINES), National University of Sciences and Technology (NUST), Islamabad, Pakistan · College of Electrical and Mechanical Engineering (CEME), NUST, Rawalpindi, Pakistan