Query-based Cross-Modal Projector Bolstering Mamba Multimodal LLM

SooHwan Eom, Jay Shim, Gwanhyeong Koo, Haebin Na, Mark Hasegawa-Johnson, Sungwoong Kim, Chang Yoo


Abstract
The Transformer’s quadratic complexity with input length imposes an unsustainable computational load on large language models (LLMs). In contrast, the Selective Scan Structured State-Space Model, or Mamba, addresses this computational challenge effectively. This paper explores a query-based cross-modal projector designed to bolster Mamba’s efficiency for vision-language modeling by compressing visual tokens based on input through the cross-attention mechanism. This innovative projector also removes the need for manually designing the 2D scan order of original image features when converting them into an input sequence for Mamba LLM. Experimental results across various vision-language understanding benchmarks show that the proposed cross-modal projector enhances Mamba-based multimodal LLMs, boosting both performance and throughput.
Anthology ID:
2024.findings-emnlp.827
Volume:
Findings of the Association for Computational Linguistics: EMNLP 2024
Month:
November
Year:
2024
Address:
Miami, Florida, USA
Editors:
Yaser Al-Onaizan, Mohit Bansal, Yun-Nung Chen
Venue:
Findings
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
14158–14167
Language:
URL:
https://aclanthology.org/2024.findings-emnlp.827
DOI:
Bibkey:
Cite (ACL):
SooHwan Eom, Jay Shim, Gwanhyeong Koo, Haebin Na, Mark Hasegawa-Johnson, Sungwoong Kim, and Chang Yoo. 2024. Query-based Cross-Modal Projector Bolstering Mamba Multimodal LLM. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 14158–14167, Miami, Florida, USA. Association for Computational Linguistics.
Cite (Informal):
Query-based Cross-Modal Projector Bolstering Mamba Multimodal LLM (Eom et al., Findings 2024)
Copy Citation:
PDF:
https://aclanthology.org/2024.findings-emnlp.827.pdf