Fine-tuning VLMs on synthetic data from TURTLEAI-Datagen improves performance on TURTLEAI, raising symbolic success rates by over 10% over base counterparts¶

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? Fine-tuning models (Pixtral-12B-TURTLE, Qwen2-VL-7B-TURTLE, Qwen2-VL-72B-TURTLE) consistently improves symbolic success rates by over 10% over their base counterparts on TURTLEAI-DS. → Chao Wen 2026 q2 i? Fine-tuning Qwen2-VL-72B on synthetic data yields roughly a 20% improvement 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

Main results table comparing fine-tuned models (trained on TURTLEAI-Train, 738k samples) against base models across DSReal, DSCraft, DSSyn, and the full dataset. For example, Qwen2-VL-72B-TURTLE reached 35.29% symbolic success on DSReal versus 11.76% for the base model; the article reports "increasing symbolic success rates by over 10%". No effect size is printed.

"Fine-tuning models (e.g., Pixtral-12B-TURTLE, Qwen2-VL-7B-TURTLE, and Qwen2-VL-72B-TURTLE) consistently improves performance on TURTLEAI-DS, increasing symbolic success rates by over 10% over their base counterparts."

Discussion¶

Related Claims¶

• 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
• Fine-tuning primarily improves alignment between code generation and visual reasoning rather than visual understanding or spatial reasoning — 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
• 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 — related
• All evaluated models struggle with spatial reasoning and precise visual replication on Turtle Graphics tasks — related