Most audio-visual language models are built and tested on short clips, which quietly avoids the thing that makes real video hard. In a long recording the sound and the picture drift in and out of agreement, evidence for an answer can be minutes apart, and what you heard early on changes how you should read what you see later.

AV-Flamingo is aimed at that longer form, jointly reasoning over audio, images and long videos, and the authors make a point of it being fully open. That matters for a class of model where the strongest systems are usually closed and therefore impossible to inspect or build on.

Their first contribution is a large collection of real-world audio-visual skills data. That ordering is telling: for long-form understanding the shortage was never architectures, it was training material that actually contains the long-range dependencies a model is supposed to learn.