Conference
Open access
2026
Enhancing Zero-Shot Visual Reasoning with Qwen2-VL on CLEVR
Overall, the results show that the baseline is fairly stable under lightweight prompt and decoding changes, however, these low-cost adjustments do not solve the model's main weakness, which still appears in spatial and compositional reasoning.
Zhong-Tian Liu
· ITM Web of Conferences · 0 citations