[论文] Beyond Modality Preference: Controlling Multimodal Arbitration under M...
研究领域: CV 作者: Zhaoying Pan, Shenyu Lu, Jun Hui Wong, Junyi Chai, Xiaoqian Wang 发布时间: 2026-11-09 arXiv: 2611.12883
论文概要
研究领域: CV 作者: Zhaoying Pan, Shenyu Lu, Jun Hui Wong, Junyi Chai, Xiaoqian Wang 发布时间: 2026-11-09 arXiv: 2611.12883
中文摘要
多模态大语言模型(MLLM)跨模态整合信息,但当模态间出现冲突时变得不可靠。在冲突输入下,MLLM 可能偏袒对任务并非最可靠的模态,这促使我们需要在推理时控制模态仲裁。我们识别出一小簇注意力头——称为仲裁头(arbitration heads)——它们对预测结果遵循视觉还是文本证据具有因果性影响。在此基础上,我们提出一种轻量级、免训练的干预方法:在推理时将这批头偏向指定的可信模态。所识别头在标定集规模变化下保持稳定,可跨设置泛化;引导它们即可在多种基准上实现模态依赖的双向有效控制。我们的发现确立了一种实用控制机制,使 MLLM 预测与任务指定的模态优先级对齐。代码:https://github.com/zhaoyingpan/modality-arbitration
原文摘要
Multimodal large language models (MLLMs) integrate information across modalities, but become unreliable when those modalities disagree. We identify a small set of attention heads, termed arbitration heads, that causally influence whether predictions follow visual or textual evidence, and introduce a lightweight, training-free intervention that biases these heads toward a designated trusted modality at inference time.
*自动采集于 2026-11-11*
#论文 #arXiv #CV #小凯