Researchers have introduced Vision2Code, a new benchmark designed to evaluate the ability of vision-language models (VLMs) to generate executable code from images. This benchmark addresses limitations of existing tools by covering 15 diverse domains, including charts, graphs, and 3D scenes, and uses a reference-code-free evaluation method. The system executes generated programs and compares their visual outputs to the source images, demonstrating higher correlation with human judgment than standard metrics. Vision2Code also aids in fine-tuning models, showing significant improvements for Qwen3.5-9B and Phi-4-MM. AI
IMPACT This benchmark could accelerate the development of more capable vision-language models for code generation tasks.
RANK_REASON The cluster describes a new academic benchmark for evaluating AI models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
