Fine-tuning primarily improves alignment between code generation and visual reasoning rather than visual understanding or spatial reasoning¶

Claim · All claims
Evidence · 1 study · causal r1 · 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 increases Qwen2-VL-72B's success rate on DSReal from 10.8% to 35.3%, mainly by reducing programming errors from 30.4% to 6.9%, while visual understanding and spatial reasoning errors remain largely unchanged. → 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? · causal · r1

Failure-type distribution analysis on DSReal comparing base and fine-tuned Qwen2-VL-72B (Figure 6). The comparison showed fine-tuning "increases the success rate from 10.8% to 35.3%", driven by reduced programming errors; visual understanding and spatial reasoning errors remained largely unchanged. No effect size is printed.

"By comparing Qwen2-VL-72B and Qwen2-VL-72B-TURTLE, we observe that fine-tuning increases the success rate from 10.8% to 35.3%, mainly by reducing programming errors (from 30.4% to 6.9%)."

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 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
• 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
• Embodied teaching actions are spatially accurate across models, but temporal action–speech alignment remains more variable — related