Tag
The paper shows that five frozen, independently trained models (four transformers and one state-space model) can be aligned into a shared representational space via Generalized Procrustes Analysis on activations alone, achieving 0.91–0.93 held-out retrieval accuracy at final layers, while reporting a negative result that routing among the models does not beat always using the strongest model.
This paper presents a method for dense latent communication between heterogeneous multi-agent systems using aligned KV-cache transformation, achieving better performance than text-based methods with lower computational costs.