VIGILant and DiCoBench tackle MLLM hallucination and evaluation

Arxiv CS.CV (computer vision)

VIGILant uses counterfactual visual alignment to reduce visual laziness and hallucinations in MLLMs. DiCoBench benchmarks multi-image fine-grained perception via differential and commonality visual cues.

MultimodalVision-language

See 4 more sources

Is Our Benchmark Enough? An Analysis of Continual Learning for MLLMs5 days agoVan-Tuan Tran, Shruthi Gowda, Merim Dzaferagic, Marco Ruffini

Spectral Evolution-Guided Token Pruning in Multimodal Large Language Models4 days agoBin Chen, Yuxiang Cai, Yadan Luo, Yi Zhang, Jianwei Yin, Zhi Chen

Mind the Heads: Topological Representation Alignment for Multimodal LLMs4 days agoDavide Caffagni, Alberto Compagnoni, Federico Melis, Sara Sarto, Pier Luigi Dovesi, Mark Granroth-Wilding, Marcella Cornia, Lorenzo Baraldi

DiCoBench: Benchmarking Multi-Image Fine-Grained Perception via Differential and Commonality Visual Cues2 days agoGeng Li, Yuxin Peng

Arxiv CS.CV (computer vision)·Xi XiaoChen LiuChih-Ting LiaoYunbei ZhangQizhen LanYuxiang WeiLin ZhaoJanet WangJianyang GuMuchao YeTianyang WangHao Xu

Jun 26, 4:00 AM