Pedestrian Attribute Recognition (PAR) plays a key role in surveillance scenarios where classical biometric traits, such as facial features, are often unavailable due to low image quality, occlusions, or variable conditions. By extracting soft biometric attributes, such as gender, clothing type, and carried objects, PAR provides essential contextual information that can support tasks like person re-identification and behavior analysis. In this work, a novel approach is proposed based on Visual Question Answering (VQA) models, which avoids the limitations of supervised learning methods by leveraging general-purpose models without the need for additional training. This extends the PAR2023-winning strategy by introducing two state-of-the-art models, PaliGemma 1 and PaliGemma 2, along with a refined set of attribute-specific questions and an innovative fusion mechanism that combines both models’ strengths. Experimental results on the PAR2025 dataset demonstrate that the proposed system surpasses previous methods, achieving a mean accuracy of 95.4% on the private set, outranking previous approaches on this task.