notesum.ai

Published at November 26

Rethinking Token Reduction in MLLMs: Towards a Unified Paradigm for Training-Free Acceleration

cs.CV

Released Date: November 26, 2024

Authors: Yuhang Han¹, Xuyang Liu², Pengxiang Ding³, Donglin Wang³, Honggang Chen², Qingsen Yan¹, Siteng Huang⁴

Aff.: ¹Northwestern Polytechnical University; ²Sichuan University; ³Westlake University; ⁴DAMO Academy, Alibaba Group

Arxiv: http://arxiv.org/abs/2411.17686v1

[Uncaptioned image]

Method	Original	Deconstructed	$\Delta$
		SQA
ToMe [3]	65.43	65.42	0.01
EViT [21]	65.21	65.18	0.03
FastV [5]	66.98	66.99	-0.01
		TextVQA
ToMe [3]	52.14	52.14	0.00
EViT [21]	51.72	51.74	-0.02
FastV [5]	52.83	52.82	0.01