ParVL Scales Multimodal LLMs by Adding Parallel Branches Over a Shared Backbone Instead of More Parameters or Longer Inference

arXiv 2608.04010·medium signalXBlueskyLinkedInCopy link

Existing MLLM scaling either grows parameters or extends sequential inference compute, and neither changes the fixed compute split between the Vision Transformer and the LLM. ParVL reuses existing ViT and LLM backbone parameters across multiple vision and language branches, instantiating each stream with branch-specific prefix parameters, trained end-to-end with full-parameter SFT on roughly 13B tokens. It beats same-recipe single-branch baselines, and the key practitioner result is that the optimal vision-versus-language compute allocation varies by task rather than having one right answer; code is at github.com/YangYangGirl/ParVL.