ParVL: Parallel Scaling and Expandable Compute Allocation for Multimodal LLMs

ParVL proposes parallel scaling for multimodal LLMs by reusing shared Vision Transformer and LLM backbone parameters across multiple vision and language branches, each differentiated through branch-specific prefix parameters. This expands inference compute without duplicating the full backbone and makes the allocation between visual encoding and language decoding adjustable. The authors report end-to-end, full-parameter supervised fine-tuning on roughly 13 billion tokens and improvements over single-branch baselines trained with the same recipe. Their experiments also indicate that the best vision-language compute allocation depends on the downstream task.