Multimodal
Multimodal models accept or produce more than one modality — text, images, audio, video. Architectures fuse encoders or use unified tokenizers across modalities.
RAG and agents increasingly mix text tools with vision/audio inputs; evaluate each modality’s failure modes separately.