---
title: "ParVL: Parallel Scaling and Expandable Compute Allocation for Multimodal LLMs"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.04010v1"
permalink: "https://feed.astralor.com/items/itm_003d0e5b0c7d1613ffde5341"
published: "2026-08-04T17:59:58.000Z"
score: 72
channel: "tech"
tags: ["ParVL","MLLMs","Parallel Scaling","Vision Transformer","Compute allocation"]
---

# ParVL: Parallel Scaling and Expandable Compute Allocation for Multimodal LLMs

## AI commentary

> The value of this paper lies in shifting the scaling perspective from 'adding parameters' to 'rearranging compute', which is especially relevant for those deploying multimodal models. If you care about deployment cost or task customization, ParVL's allocatable mechanism may inspire more efficient designs.

## At a glance

Existing scaling strategies for multimodal large language models typically expand model parameters or sequential inference computation, incurring high memory or latency overhead, and fail to adjust the rigid compute allocation between vision transformers and language models. ParVL introduces a parallel scaling framework that reuses existing ViT and LLM backbone parameters across multiple vision and language branches, enabling flexible compute allocation between modalities under a fixed backbone budget. After full-parameter supervised fine-tuning on roughly 13B tokens, ParVL outperforms same-recipe single-branch baselines in overall multimodal performance, with optimal vision-language allocation varying across tasks. Code is available on GitHub.

## Key points

- 提出ParVL框架，通过共享骨干复用实现并行计算扩展，而非增加参数或串行深度
- 系统研究固定参数预算下视觉与语言模态之间的计算分配权衡
- 在约130亿token上全参数微调，整体性能优于单分支基线，且最优分配因任务变化
- 开源代码已发布

## Full text

Existing scaling strategies for Multimodal Large Language Models (MLLMs) typically expand either model parameters or sequential inference computation, incurring substantial memory or latency overhead. More importantly, most existing methods fail to alter the rigid, fixed computation allocation between the Vision Transformer and the Large Language Model components, limiting task-specific optimization. To address this, we introduce the Parallel Vision-Language (ParVL) scaling framework for MLLMs, which scales parallel computation by reusing the existing ViT and LLM backbone parameters across multiple vision and language branches. This framework raises a central question: given a fixed backbone parameter budget, how should additional shared-backbone computation be allocated between the vision and language modalities? We instantiate each parallel computational stream with branch-specific prefix parameters over a shared backbone, and train the entire model end-to-end via full-parameter supervised fine-tuning on roughly 13B tokens. We systematically study the computation-allocation trade-off between the ViT encoder and LLM decoder. ParVL improves overall multimodal performance over same-recipe single-branch baselines, and the best evaluated vision--language allocation varies across tasks. Code is available at https://github.com/YangYangGirl/ParVL.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_003d0e5b0c7d1613ffde5341