LVLM-CL: Make Large Vision-Language Models Work Better under Continual Learning Settings

Baiqiao Yin

Keywords: Continual Learning; Large vision-language model

Abstract:

The development of Large Vision-Language Models
(LVLMs) is striving to catch up with the success of Large
Language Models (LLMs), yet it faces more challenges to
be resolved. When finetuning LVLMs with user-specific
data in the practical use, the pretrained weights would face
the problems of forgetting and performance degradation.
So it is important to improve LVLM’s performance under
the continual learning settings. Some existing CL methods
have explored continual learning on VLM. However, the
continual learning settings they have proposed couldn’t
be adopted to LVLMs smoothly because the training and
finetuning process of LVLMs need amount of data while
previous VLM continual learning settings built on limited
data and different model architectures. In this work, we first
devise a task-specific continual learning setting especially
for LVLMs by classifying the instruction tuning data
for the second finetune process of LVLMs into several
different tasks. Mimicking the process of finetuning with
user-specific task data, we found that the performance of
LVLMs would decline without any modules designed for
continual learning settings. So we present LVLM-CL, a
novel approach capable of continual learning settings for
large vision-language models when finetuning with different
kinds of tasks. Specifically, our LVLM-CL consists of a
text feature based prompt that are different between tasks
to keep the special feature of different tasks. To meet the
setting of continual learning, we also design a memory bank
which storage previous trained tasks which helps LVLMs
apply knowledge to unfamiliar combinations. Extensive case
studies and quantitative evaluations show LVLM-CL has
strong capability in understanding the pivotal features of
different tasks and emerges impressive memory capabilities
under the continual learning settings. This work fosters the
advancements of LVLMs by enabling them to support better
continual finetuning toward practical use in the real world.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3217