Current vision-language models struggle on education-oriented Turtle Graphics tasks, with even top models far below reliable performance on real-world tasks for grades 3-6¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? On the real-world dataset DSReal, the top-performing model o3 achieves only a 40.2% symbolic success rate on tasks designed for students in grades 3-6. → Chao Wen 2026 q2 i? On the full TURTLEAI-DS dataset, the highest base-model symbolic success rate is 15.9% (o4-mini), and most models fall below 30% on real-world tasks. → Chao Wen 2026

Evidence¶

Chao Wen 2026¶

Chao Wen, Jacqueline Staub, and Adish Singla. (2026). TURTLEAI: Benchmarking Multimodal Models for Visual Programming in Turtle Graphics. https://arxiv.org/abs/2606.03626 doi:10.48550/arxiv.2606.03626

q2 · i? · design · r2

Benchmark evaluation of 20+ VLMs on TURTLEAI-DSReal (102 real-world tasks), reporting symbolic and embedding-based success rates with greedy decoding. The evaluation found "only a 40.2% symbolic success rate" for the best model, o3; on the full dataset the best base model reached 15.9% (o4-mini) and Qwen2.5-VL 6.80%. No effect size is reported.

"On the real-world dataset DSReal containing tasks designed for students in grades 3-6, even the top-performing model, o3, achieves only a 40.2% symbolic success rate, suggesting current VLMs are still far from ready for classroom deployment."

Discussion¶

Related Claims¶

• Fine-tuning primarily improves alignment between code generation and visual reasoning rather than visual understanding or spatial reasoning — related
• All evaluated models struggle with spatial reasoning and precise visual replication on Turtle Graphics tasks — related
• Fine-tuning VLMs on synthetic data from TURTLEAI-Datagen improves performance on TURTLEAI, raising symbolic success rates by over 10% over base counterparts — related
• Fine-tuning on clean synthetic data yields limited gains on hand-drawn tasks, indicating limited robustness to out-of-distribution hand-drawn inputs — related
• Model performance varies by task category, with Basic Geometry easiest and Spiral consistently the most challenging category — related
• Fine-tuning is associated with more compact generated code, with the fine-tuned Qwen2-VL-72B maintaining a length ratio near 1.0 while the base model's failed outputs become substantially longer — related