Multi-agent systems rely on communication to accomplish tasks, and latent communication reduces text generation and re‑encoding by passing continuous representations. Existing latent‑communication methods assume fixed model combinations, causing interface failures when agents are replaced. Our pilot experiments confirm this issue and highlight the need for interface reuse after model replacement.
LatBridge connects agent models via a shared Bridge and separate interfaces under a common latent‑communication protocol. After a model swap, unchanged interfaces are retained, matching low‑rank interfaces are selected, and they are combined through the Bridge. Only the missing interfaces are trained while all other models stay frozen. The Bridge processes source representations in a basic mode and constructs target inputs in an accelerated mode, eliminating autoregressive generation.
Experiments on eight small‑model configurations show mean basic‑mode accuracies of 14.8% on MATH and 27.3% on ALFWorld for next‑action prediction. Analyses reveal that interface compatibility, low‑rank projection, and the Bridge’s aggregation are the main factors supporting task accuracy and communication efficiency.
Review