Reproduce “DeepVIS: Bridging Natural Language and Data Visualization Through Step-wise Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhihao Shuai, Boyan Li, Siyu Yan, Yuyu Luo, Weikai Yang
DeepVIS enables transparent and interactively steerable data visualization from natural language queries by employing a five-stage Chain-of-Thought reasoning process. This approach, which fine-tunes a small-scale LLM on an augmented dataset, achieves state-of-the-art 77.16% overall accuracy in NL2VIS, outperforming larger general-purpose LLMs while improving user comprehension and refinement efficiency.
Key insights
- Decomposing the NL2VIS task into a structured, step-wise Chain-of-Thought reasoning process substantially enhances LLM performance and output transparency.
- Augmenting LLM inputs with database schema descriptions and value sampling is critical for accurate data retrieval and conditional logic in visualization generation.
- Interactive interfaces that expose and allow refinement of intermediate reasoning steps significantly improve a user's ability to identify errors and efficiently steer the AI's output.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.