ASCII Generation Understanding Benchmark

Unlocking the Latent Canvas: Eliciting and Benchmarking Symbolic Visual Expression in LLMs

We introduce SVE-ASCII: an open 7B model (SVE-ASCII-7B), a reproducible dataset, and a standardized benchmark for symbolic visual expression via ASCII art.

📄 Paper arXiv 💻 Code GitHub 🤗 Model HF 📚 Dataset HF 📊 Benchmark HF

What is released?

Model: SVE-ASCII-7B · Dataset: SVE-ASCII-Dataset · Benchmark: SVE-ASCII-Bench (Generation + Understanding)

Prefer details? See the GitHub README.

Qualitative Results

SVE-ASCII Benchmark

SVE-ASCII-Bench covers two tasks: Generation (text → ASCII art), Understanding (ASCII art → description / label), Generation uses SA / IF / SC / SL / CE and a weighted composite score on Recall (in-distribution) and Generalization (OOD) splits.

Quick Start

# Clone & download benchmark from HF
cd SVE-ASCII/benchmark

# Generation
python eval_generation.py

# Understanding
python eval_understanding.py

# Understanding-Selection
python eval_understanding_selection.py

Set MODEL_TYPE, API keys, etc. See benchmark/README.md.

Resources

Links

🤗

Model SVE-ASCII-7B

Open ↗

📚

Dataset SVE-ASCII-Dataset

Open ↗

📊

Benchmark SVE-ASCII-Bench

Open ↗

💻

Code evaluation + synthesis

Repo ↗

Citation

BibTeX

@article{yourname2026sveascii,
  title={Unlocking the Latent Canvas: Eliciting and Benchmarking Symbolic Visual Expression in LLMs},
  author={Your Name and Others},
  journal={arXiv preprint arXiv:XXXX.XXXXX},
  year={2026}
}