![[2602.15382] The Vision Wormhole: Latent-Space Communication in Heterogeneous Multi-Agent Systems](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.15382] The Vision Wormhole: Latent-Space Communication in Heterogeneous Multi-Agent Systems
Summary
The paper introduces the Vision Wormhole, a framework for enabling efficient latent-space communication in heterogeneous multi-agent systems, improving collaborative reasoning without relying on text-based communication.
Why It Matters
As multi-agent systems increasingly utilize large language models, the inefficiencies of traditional text communication hinder performance. The Vision Wormhole offers a scalable, model-agnostic solution that enhances communication speed and fidelity, which is crucial for advancing AI collaboration in diverse applications.
Key Takeaways
- The Vision Wormhole framework enables text-free communication between heterogeneous models.
- It reduces communication overhead and improves efficiency in multi-agent systems.
- The framework employs a Universal Visual Codec for mapping reasoning traces into a shared latent space.
- Experimental results show reduced processing time while maintaining reasoning fidelity.
- The approach simplifies pairwise alignment complexity significantly.
Computer Science > Computation and Language arXiv:2602.15382 (cs) [Submitted on 17 Feb 2026] Title:The Vision Wormhole: Latent-Space Communication in Heterogeneous Multi-Agent Systems Authors:Xiaoze Liu, Ruowang Zhang, Weichen Yu, Siheng Xiong, Liu He, Feijie Wu, Hoin Jung, Matt Fredrikson, Xiaoqian Wang, Jing Gao View a PDF of the paper titled The Vision Wormhole: Latent-Space Communication in Heterogeneous Multi-Agent Systems, by Xiaoze Liu and 9 other authors View PDF HTML (experimental) Abstract:Multi-Agent Systems (MAS) powered by Large Language Models have unlocked advanced collaborative reasoning, yet they remain shackled by the inefficiency of discrete text communication, which imposes significant runtime overhead and information quantization loss. While latent state transfer offers a high-bandwidth alternative, existing approaches either assume homogeneous sender-receiver architectures or rely on pair-specific learned translators, limiting scalability and modularity across diverse model families with disjoint manifolds. In this work, we propose the Vision Wormhole, a novel framework that repurposes the visual interface of Vision-Language Models (VLMs) to enable model-agnostic, text-free communication. By introducing a Universal Visual Codec, we map heterogeneous reasoning traces into a shared continuous latent space and inject them directly into the receiver's visual pathway, effectively treating the vision encoder as a universal port for inter-agent telepathy. Ou...