ParVL: Parallel Scaling and Expandable Compute Allocation for Multimodal LLMs
ParVL adds parallel vision and language branches while reusing the same ViT and LLM backbone weights.
The paper argues this can scale multimodal inference without simply adding parameters or lengthening sequential compute. Each branch uses its own prefix parameters over the shared backbone, with the model trained end to end on roughly 13B tokens. The authors report gains over single-branch baselines, and find that the best split of extra compute between vision and language changes by task. ArXiv · AI/CL/LG's note
The paper argues this can scale multimodal inference without simply adding parameters or lengthening sequential compute. Each branch uses its own prefix parameters over the shared backbone, with the model trained end to end on roughly 13B tokens. The authors report gains over single-branch baselines, and find that the best split of extra compute between vision and language changes by task. ArXiv · AI/CL/LG's note
score 5