We validate a recipe for large-scale, multi-task RL for visual reasoning, here are the key efforts:
🎯Task-specific verifiable rewards beyond VLM
⏩1-step delay + shared-prompt waves overlap CPU scoring with GPU replay
⚡FSDP2/HSDP/TP + Liger/FA3/torch.compile
🌊Support SDE/CPS
02:44
💡Videos and images should not be treated merely as inputs to understand or outputs to render. They can be first-class substrates for problem solving beyond language.
🎉Introducing VBVR-Pro: an infrastructure suite for visual intelligence: trainable, verifiable, RL-optimizable,
