Splash utilizes mask-isolated tactile alignment learning to integrate touch into multimodal LLMs without degrading existing vision-language reasoning. The method partitions parameters into a frozen critical subspace for visual knowledge and an updated dormant subspace for tactile sensing.
HOW THIS AFFECTS YOU
●
researcherYou can integrate new sensory modalities into compact MLLMs without catastrophic forgetting.
●
designerThis could lead to more physically aware AI interactions in robotics and AR.