VLsI: Verbalized Layers-to-Interactions from Large to Small Vision Language Models

TL;DR

VLsI introduces a layer-wise distillation approach with verbalizers to efficiently scale small vision-language models, achieving significant benchmark improvements without increasing model size or complexity.

Contribution

The paper presents VLsI, a novel layer-wise distillation method with verbalizers for efficient vision-language model scaling, outperforming larger models without additional scaling.

Findings

Achieves 11.0% and 17.4% improvements on benchmarks for 2B and 7B models.

Validates effectiveness across ten vision-language benchmarks.

Reduces computational costs while maintaining high accuracy.

Abstract

The recent surge in high-quality visual instruction tuning samples from closed-source vision-language models (VLMs) such as GPT-4V has accelerated the release of open-source VLMs across various model sizes. However, scaling VLMs to improve performance using larger models brings significant computational challenges, especially for deployment on resource-constrained devices like mobile platforms and robots. To address this, we propose VLsI: Verbalized Layers-to-Interactions, a new VLM family in 2B and 7B model sizes, which prioritizes efficiency without compromising accuracy. VLsI leverages a unique, layer-wise distillation process, introducing intermediate "verbalizers" that map features from each layer to natural language space, allowing smaller VLMs to flexibly align with the reasoning processes of larger VLMs. This approach mitigates the training instability often encountered in output imitation and goes beyond typical final-layer tuning by aligning the small VLMs' layer-wise progression with that of the large ones. We validate VLsI across ten challenging vision-language benchmarks, achieving notable performance gains (11.0% for 2B and 17.4% for 7B) over GPT-4V without the need for model scaling, merging, or architectural changes.