ASCII Diagram Benchmark for VLMs Released
Key point
ASCIITermDraw-Bench, which evaluates VLMs' ability to generate and edit text-based diagrams, has been released.
Details
ASCIITermDraw-Bench has been introduced to evaluate whether VLMs (Vision Language Models) can go beyond simple text descriptions to generate and edit accurately structured diagrams using ASCII characters.
Unlike existing coding- or math-focused evaluations, this benchmark measures the ability to implement visual structures with complex layouts in text form. The main evaluation areas are as follows:
- Basic boxes and layouts
- Network topologies
- Software architecture diagrams
- Image-conditioned diagram editing (the ability to modify specific parts while preserving the existing structure)
Evaluation consists of a Structural Score that measures structural accuracy and a Semantic Score via an LLM judge. On the current leaderboard, Gemma-4-31B-IT records the highest performance at 73.8%, followed by Qwen3.7-Plus (70.2%).
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.