cross-modal transfer from vision RL to text

vision-text early fusion

Joint pretraining is the enabling condition that lets multimodal RL feed back into text capability rather than only adding vision as a downstream feature.

1 source videos