VisCoder2: Building Multi-Language Visualization Coding Agents

2025-10-24

AI Summary

This paper addresses three critical bottlenecks of existing large language model–driven visualization coding agents in practical workflows: narrow programming language coverage, unreliable code execution, and absence of iterative refinement capability. To tackle these, we propose a systematic solution comprising two core components: (1) the construction of VisCode-Multi-679K—a large-scale, multilingual, multi-turn dialogue dataset—and VisPlotBench—a novel executable evaluation benchmark; and (2) the VisCoder2 model family, supporting 12 programming languages, integrating large-scale supervised training, multi-turn self-debugging, and cross-language execution verification. Experimental results show that the 32B VisCoder2 achieves an execution success rate of 82.4%, significantly outperforming open-source baselines and approaching GPT-4.1 performance—particularly excelling in symbol-sensitive or compiler-dependent languages.

Technology Category

• Computer Vision: Language and Vision
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Large language models (LLMs) have recently enabled coding agents capable of generating, executing, and revising visualization code. However, existing models often fail in practical workflows due to limited language coverage, unreliable execution, and lack of iterative correction mechanisms. Progress has been constrained by narrow datasets and benchmarks that emphasize single-round generation and single-language tasks. To address these challenges, we introduce three complementary resources for advancing visualization coding agents. VisCode-Multi-679K is a large-scale, supervised dataset containing 679K validated and executable visualization samples with multi-turn correction dialogues across 12 programming languages. VisPlotBench is a benchmark for systematic evaluation, featuring executable tasks, rendered outputs, and protocols for both initial generation and multi-round self-debug. Finally, we present VisCoder2, a family of multi-language visualization models trained on VisCode-Multi-679K. Experiments show that VisCoder2 significantly outperforms strong open-source baselines and approaches the performance of proprietary models like GPT-4.1, with further gains from iterative self-debug, reaching 82.4% overall execution pass rate at the 32B scale, particularly in symbolic or compiler-dependent languages.

Problem

• Building multi-language visualization coding agents
• Addressing limited language coverage and unreliable execution
• Overcoming narrow datasets and single-round generation constraints

Innovation

• Multi-language dataset with 679K executable visualization samples
• Benchmark for systematic evaluation of generation and debugging
• Family of models trained on large-scale multi-language dataset