Ankan Deria

A single visual world model can now execute complex procedural tasks end-to-end, adapting its generation to the actual outcome of each predicted action.

Expensive end-to-end omni pre-training degrades foundational reasoning and vision, yet simply routing timestamped ASR transcripts into a frozen VLM matches or beats native omni architectures across 56 benchmarks.

Forget choosing just one vision encoder – fusing CLIP and DINO representations unlocks a significant performance boost in vision-language tasks.