Investigating the Catastrophic Forgetting in Multimodal Large Language Model Fine-Tuning

Yuexiang Zhai, Shengbang Tong, Xiao Li, Mu Cai, Qing Qu, Yong Jae Lee, Yi Ma
Conference on Parsimony and Learning, PMLR 234:202-227, 2024.

Abstract

Following the success of GPT4, there has been a surge in interest in multimodal large language model (MLLM) research. This line of research focuses on developing general-purpose LLMs through fine-tuning pre-trained LLMs and vision models. However, catastrophic forgetting, a notorious phenomenon where the fine-tuned model fails to retain similar performance compared to the pre-trained model, still remains an inherited problem in multimodal LLMs (MLLM). In this paper, we introduce EMT: Evaluating MulTimodality for evaluating the catastrophic forgetting in MLLMs, by treating each MLLM as an image classifier. We first apply EMT to evaluate several open-source fine-tuned MLLMs and we discover that almost all evaluated MLLMs fail to retain the same performance levels as their vision encoders on standard image classification tasks. Moreover, we continue fine-tuning LLaVA, an MLLM and utilize EMT to assess performance throughout the fine-tuning. Interestingly, our results suggest that early-stage fine-tuning on an image dataset improves performance across other image datasets, by enhancing the alignment of text and language features. However, as fine-tuning proceeds, the MLLMs begin to hallucinate, resulting in a significant loss of generalizability, even when the image encoder remains frozen. Our results suggest that MLLMs have yet to demonstrate performance on par with their vision models on standard image classification tasks and the current MLLM fine-tuning procedure still has room for improvement.

Cite this Paper

BibTeX
@InProceedings{pmlr-v234-zhai24a, title = {Investigating the Catastrophic Forgetting in Multimodal Large Language Model Fine-Tuning}, author = {Zhai, Yuexiang and Tong, Shengbang and Li, Xiao and Cai, Mu and Qu, Qing and Lee, Yong Jae and Ma, Yi}, booktitle = {Conference on Parsimony and Learning}, pages = {202--227}, year = {2024}, editor = {Chi, Yuejie and Dziugaite, Gintare Karolina and Qu, Qing and Wang, Atlas Wang and Zhu, Zhihui}, volume = {234}, series = {Proceedings of Machine Learning Research}, month = {03--06 Jan}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v234/zhai24a/zhai24a.pdf}, url = {https://proceedings.mlr.press/v234/zhai24a.html}, abstract = {Following the success of GPT4, there has been a surge in interest in multimodal large language model (MLLM) research. This line of research focuses on developing general-purpose LLMs through fine-tuning pre-trained LLMs and vision models. However, catastrophic forgetting, a notorious phenomenon where the fine-tuned model fails to retain similar performance compared to the pre-trained model, still remains an inherited problem in multimodal LLMs (MLLM). In this paper, we introduce EMT: Evaluating MulTimodality for evaluating the catastrophic forgetting in MLLMs, by treating each MLLM as an image classifier. We first apply EMT to evaluate several open-source fine-tuned MLLMs and we discover that almost all evaluated MLLMs fail to retain the same performance levels as their vision encoders on standard image classification tasks. Moreover, we continue fine-tuning LLaVA, an MLLM and utilize EMT to assess performance throughout the fine-tuning. Interestingly, our results suggest that early-stage fine-tuning on an image dataset improves performance across other image datasets, by enhancing the alignment of text and language features. However, as fine-tuning proceeds, the MLLMs begin to hallucinate, resulting in a significant loss of generalizability, even when the image encoder remains frozen. Our results suggest that MLLMs have yet to demonstrate performance on par with their vision models on standard image classification tasks and the current MLLM fine-tuning procedure still has room for improvement.} }
Endnote
%0 Conference Paper %T Investigating the Catastrophic Forgetting in Multimodal Large Language Model Fine-Tuning %A Yuexiang Zhai %A Shengbang Tong %A Xiao Li %A Mu Cai %A Qing Qu %A Yong Jae Lee %A Yi Ma %B Conference on Parsimony and Learning %C Proceedings of Machine Learning Research %D 2024 %E Yuejie Chi %E Gintare Karolina Dziugaite %E Qing Qu %E Atlas Wang Wang %E Zhihui Zhu %F pmlr-v234-zhai24a %I PMLR %P 202--227 %U https://proceedings.mlr.press/v234/zhai24a.html %V 234 %X Following the success of GPT4, there has been a surge in interest in multimodal large language model (MLLM) research. This line of research focuses on developing general-purpose LLMs through fine-tuning pre-trained LLMs and vision models. However, catastrophic forgetting, a notorious phenomenon where the fine-tuned model fails to retain similar performance compared to the pre-trained model, still remains an inherited problem in multimodal LLMs (MLLM). In this paper, we introduce EMT: Evaluating MulTimodality for evaluating the catastrophic forgetting in MLLMs, by treating each MLLM as an image classifier. We first apply EMT to evaluate several open-source fine-tuned MLLMs and we discover that almost all evaluated MLLMs fail to retain the same performance levels as their vision encoders on standard image classification tasks. Moreover, we continue fine-tuning LLaVA, an MLLM and utilize EMT to assess performance throughout the fine-tuning. Interestingly, our results suggest that early-stage fine-tuning on an image dataset improves performance across other image datasets, by enhancing the alignment of text and language features. However, as fine-tuning proceeds, the MLLMs begin to hallucinate, resulting in a significant loss of generalizability, even when the image encoder remains frozen. Our results suggest that MLLMs have yet to demonstrate performance on par with their vision models on standard image classification tasks and the current MLLM fine-tuning procedure still has room for improvement.
APA
Zhai, Y., Tong, S., Li, X., Cai, M., Qu, Q., Lee, Y.J. & Ma, Y.. (2024). Investigating the Catastrophic Forgetting in Multimodal Large Language Model Fine-Tuning. Conference on Parsimony and Learning, in Proceedings of Machine Learning Research 234:202-227 Available from https://proceedings.mlr.press/v234/zhai24a.html.

Related Material