Recognizing autism-related behaviors from real-world data is a crucial step toward supporting early diagnosis and intervention. However, existing methods are often limited to single-modality inputs or rely on computationally intensive multimodal architectures, hindering practical deployment. To address these challenges, this paper presents a lightweight audio-visual framework based on the AV-ASD dataset. The approach combines TimeSformer for efficient video-based spatio-temporal representation and Wav2Vec2 for audio encoding, integrated via a Transformer-based fusion module to capture temporal and cross-modal dependencies. The proposed model achieves strong performance across key behavior categories, demonstrating the potential of efficient multimodal learning for autism behavior analysis.