具身智能观察

Beyond Multimodal Alignment: Certifying Physical Language through Response Substitution and Ordered Execution

技术动态

来源:arXiv cs.RO发布时间待核实

arXiv:2608.19492v1 Announce Type: cross Abstract: World models increasingly treat compact multimodal representations as interfaces between perception and physical interaction, yet existing probes do not establish whether different sensors carry the same executable meaning or whether that meaning survives a new action composition.

Beyond Multimodal Alignment: Certifying Physical Language through Response Substitution and Ordered Execution | 具身智能观察