Please note: This master’s thesis presentation will take place online.
Yuansheng Ni, Master’s candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Wenhu Chen
Large language models have recently improved at general-purpose code generation, but they remain unreliable for visualization tasks, where correctness depends on aligning code, data semantics, and rendered outputs. Visualization coding requires execution grounding and semantic coherence that syntax alone cannot guarantee. This thesis investigates how to build visualization coding agents that can generate plotting code, execute it, and iteratively refine it using feedback. The first part of the work focuses on Python-based visualization and introduces VisCode-200K, a dataset combining validated plotting examples with multi-turn correction dialogues. Using this resource, we develop VisCoder, a model that leverages runtime feedback to improve executable and visually accurate plot generation, achieving strong performance on PandasPlotBench relative to open-source baselines.
Building on these findings, the thesis extends visualization coding to a multi-language setting. We construct VisCode-Multi-679K, a large instruction-tuning corpus spanning 12 visualization languages, and develop VisPlotBench, an executable benchmark supporting both initial code generation and multi-round self-debugging. Using these resources, we train VisCoder2, a family of models that demonstrates substantial gains over open-source systems and approaches the performance of proprietary models, particularly in symbolic or compiler-dependent languages. Together, these components form a unified framework and resource suite for reliable, feedback-aligned visualization coding agents capable of supporting diverse real-world analytical workflows.
Attend this master’s thesis presentation virtually on MS Teams.