StoryReasoning Dataset: Using Chain-of-Thought for Scene Understanding and Grounded Story Generation (2505.10292v1)

Abstract: Visual storytelling systems struggle to maintain character identity across frames and link actions to appropriate subjects, frequently leading to referential hallucinations. These issues can be addressed through grounding of characters, objects, and other entities on the visual elements. We propose StoryReasoning, a dataset containing 4,178 stories derived from 52,016 movie images, with both structured scene analyses and grounded stories. Each story maintains character and object consistency across frames while explicitly modeling multi-frame relationships through structured tabular representations. Our approach features cross-frame object re-identification using visual similarity and face recognition, chain-of-thought reasoning for explicit narrative modeling, and a grounding scheme that links textual elements to visual entities across multiple frames. We establish baseline performance by fine-tuning Qwen2.5-VL 7B, creating Qwen Storyteller, which performs end-to-end object detection, re-identification, and landmark detection while maintaining consistent object references throughout the story. Evaluation demonstrates a reduction from 4.06 to 3.56 (-12.3%) hallucinations on average per story when compared to a non-fine-tuned model.

Overview of NeurIPS 2025 Formatting Instructions

The document, authored by David S. Hippocampus from Cranberry-Lemon University, offers comprehensive guidance for preparing and submitting papers to the NeurIPS 2025 conference. The instructions are detailed to ensure compliance with the formatting standards required by the NeurIPS committee. While focusing primarily on the technical aspects of paper formatting, it addresses submission style, retrieval of style files, general formatting, headings, citations, figures, tables, equations, PDF preparation, and the paper checklist.

Submission and Formatting Guidelines

Authors are required to adhere to specific formatting rules which include the usage of NeurIPS LaTeX style files. The maximum permitted length for papers is nine pages, excluding references and appendices. Authors must utilize the updated LaTeX style file (neurips_2025.sty) and refrain from using deprecated versions or altering the style parameters. The guideline also emphasizes the importance of maintaining consistency in citation style and addresses how best to manage situations where natbib options, commonly used for citations, clash with other loaded packages.

Figures, Tables, and Equations

Handling artwork and data representation through figures, tables, and mathematical equations is given thorough attention. Authors are guided to maintain high quality and legibility of figures and advised to use color judiciously for clarity in both colored and black-and-white prints. Similarly, tables should adhere to a standardized format suggested by the document and utilize the booktabs package for professional-grade typesetting without vertical lines.

Typefaces and Technical Compliance

The document discusses the importance of using Type 1 or Embedded TrueType fonts to avoid technical setbacks in PDF submissions. It includes actionable advice for checking font types through standard PDF readers and tools on Linux-based systems. The instructions stress the use of pdflatex for generating compliant submissions and provide guidelines on how to ensure uniform margins, particularly when including graphics.

Ethical Conduct and Broader Impact

Of particular note is the inclusion of the "NeurIPS Paper Checklist," which ensures ethical and responsible machine learning research practices. This checklist mandates authors to discuss implications concerning societal impact, reproducibility, limitations, biases, and ethical adherence to the NeurIPS code of ethics. Transparency regarding the dataset and materials used in research—including licenses and terms of use—is also required. Authors are urged to approach potential ethical concerns with due diligence, especially when releasing data or models susceptible to misuse.

Theoretical and Experimental Considerations

Authors should ensure that theoretical contributions include comprehensive proofs and clearly stated assumptions, while experimental results should be fully reproducible. The checklist questions emphasize the necessity for standard reporting practices such as declaring computational resources, detailing experimental settings, including error bars for statistical significance, and ensuring open access to data and code whenever possible.

Implications and Future Directions

The directives outlined serve as a foundational framework for the submission process to one of the most prestigious conferences in AI research. Adherence to these standards can have significant practical implications, ensuring that submissions are uniformly assessed and increasing the likelihood of acceptance. The guidelines and ethical considerations aim to bolster the integrity and quality of AI research disseminated through the NeurIPS platform, promoting reproducibility and transparency while guiding future submissions towards responsible practices in AI research.