IC-277Open-source VLMs show a clear scaling trend in both average accuracy and reasoning robustness on DynaMath, with larger models performing substantially better

Chengke Zou, Xingang Guo, Rui Yang, Junyu Zhang, Bin Hu, Huan Zhang

SourceDynaMath: A Dynamic Visual Benchmark for Evaluating Mathematical Reasoning Robustness of Vision Language Models

Among open-source models, increasing parameter size consistently improves both average-case accuracy and reasoning robustness. Qwen2-VL improves from 42.1% (7B) to 55.1% (72B) average accuracy, and InternVL2 improves from 39.7% (8B) to 54.0% (76B). The paper notes that closed-source models demonstrate greater robustness than open-source models overall, but within the open-source group, reasoning robustness scales with model size. Qwen2-VL-72B and InternVL2-76B outperform Gemini Pro 1.5 on reasoning robustness.

Evidence
correlational
Key metric
Qwen2-VL: 42.1% (7B) to 55.1% (72B) average accuracy; InternVL2: 39.7% (8B) to 54.0% (76B) average accuracy. Reasoning robustness (RR): Qwen2-VL-72B 51.8%, InternVL2-76B 45.8%, Gemini Pro 1.5 44.5%.
Caveat
The scaling trend is observed only within two model families (Qwen2-VL and InternVL2) and only up to 76B parameters; it is not established whether the trend continues beyond this range.
Model
Qwen2-VL Qwen2-VL-7B, Qwen2-VL-72B, InternVL2 InternVL2-8B, InternVL2-26B, InternVL2-40B, InternVL2-76B, Gemini 1.5 / Gemini Pro 1.5
Concepts
Scale-dependent behaviour
Datasets
DynaMath [eval]
Related findings
IC-276, IC-278
Extraction
automatic-extraction