Revision History

Bring the multimodal model survey current through July 2026: add native multimodal pretraining versus the adapter pattern, early-versus-late fusion at pretraining scale with the conflicting industry and research definitions, unified understand-and-generate designs, audio and video as first-class modalities with streaming and long-video tokenisation, benchmarks answerable without the image, and safety exposure from image and audio inputs

Correction: remove unsupported product, benchmark, architecture, and compute claims; rebuild the overview around sourced modality definitions, architectures, training, evaluation, failure modes, and reporting practice.

Added 6 contextual internal links

Auto-saved before update

Auto-saved before update

Edit details were not recorded

Initial import from MediaWiki