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¶

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 DSReal, Qwen2-VL-72B's outputs become substantially longer in failure cases (length ratio 4.01) than successful cases (1.52), while the fine-tuned Qwen2-VL-72B-TURTLE maintains a length ratio close to 1.0 across all cases. → 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

Code length analysis on DSReal computing length ratio (model-generated lines divided by reference-solution lines, comments and blank lines removed) for GPT-4o, Qwen2-VL-72B, and Qwen2-VL-72B-TURTLE. The article reports the base model's failure-case ratio of 4.01 versus 1.52 on successes, and that the fine-tuned model "maintains a length ratio close to 1.0 across all cases". No effect size is printed.

"Qwen2-VL-72B exhibits a moderate length ratio on successful cases (ratio = 1.52), but its outputs become substantially longer in failure cases (ratio = 4.01), indicating a substantial increase in code verbosity upon failure."

Discussion¶

Related Claims¶

• Fine-tuning primarily improves alignment between code generation and visual reasoning rather than visual understanding or spatial reasoning — 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
• All evaluated models struggle with spatial reasoning and precise visual replication on Turtle Graphics tasks — 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
• Fine-tuning on clean synthetic data yields limited gains on hand-drawn tasks, indicating limited robustness to out-of-distribution hand-drawn inputs — related