Vision Wormhole: Latent-Space Communication in Heterogeneous Multi-Agent Systems
The Vision Wormhole framework solves communication bottlenecks in multi-agent systems using Large Language Models by mapping reasoning traces into a shared continuous visual space. It enables cross-architecture latent state transfer between different VLMs without needing specific translators. This allows agents to collaborate efficiently by exchanging compact, universal tokens, significantly reducing runtime overhead and information loss.