Unbundling Multimodal Vision Encoding
Unbundling Multimodal Vision Encoding
Control experiments show that multimodal language models do not require pre-trained visual encoders, but benefit from dedicated visual capabilities trained jointly with the backbone, and this capability should be integrated early rather than being deployed before fusion. The study pre-trained all models from scratch on public data and compared them on visual question answering and language benchmarks with matched computational costs, finding that visual priors are not necessary; encoders trained jointly from scratch are as effective as pre-trained encoders and do not require a separate training stage.
Published on Nov 12
Authors:
,
,
,
,
,
,
,
,
,
Abstract
Controlled experiments show that multimodal language models do not need a pretrained vision encoder, but benefit from dedicated vision capacity trained jointly with the backbone and integrated early rather than before fusion.
How multimodal language models should encode visual input remains up for debate. Most models use a pretrained vision encoder, while some recent recipes train the encoder from scratch with the language model or remove it entirely. Vision encoders bundle together three design choices: a visual prior from upstream pretraining, dedicated vision capacity, and pre-fusion placement of that capacity (i.e., before the multimodal backbone). Because published models also differ in data, architecture, and compute, comparing them cannot isolate the effect of each choice. We disentangle these choices via a controlled framework where we pretrain all models from scratch on common data and compare them on visual question answering and language benchmarks at matched lifetime compute. We find that the visual prior is not necessary; an encoder trained jointly from scratch is on par with a pretrained one, without a separate training stage. What matters instead is dedicated vision capacity: an encoder-free model trails on visual understanding even when its backbone is enlarged by the encoder's parameter count. Yet that capacity need not precede multimodal fusion, as vision-only modules added in-fusion to early backbone layers perform comparably to a pre-fusion encoder and can lower time to first token. These results also hold across backbone sizes, when encoder pretraining is treated as free, and when starting from a pretrained language model. Overall, our findings favor dedicated vision capacity trained jointly with the backbone and placed early within it.
View arXiv page View PDF Project page Add to collection
Get this paper in your agent:
hf papers read 2611.17334
Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash
Models citing this paper 0
No model linking this paper
Cite arxiv.org/abs/2611.17334 in a model README.md to link it from this page.
Datasets citing this paper 0
No dataset linking this paper
Cite arxiv.org/abs/2611.17334 in a dataset README.md to link it from this page.
Spaces citing this paper 0
No Space linking this paper
Cite arxiv.org/abs/2611.17334 in a Space README.md to link it from this page.
Collections including this paper 2
Source: Hugging Face Daily Papers · huggingface.co