Search papers, labs, and topics across Lattice.
Affiliation:
1
0
3
2
Multimodal RL models trained without visual input can still achieve high accuracy on visual reasoning tasks when images are present at inference, but prolonged visual training can paradoxically erode grounding.
