Latent Communication Between Language Model Agents: Channels, Alignment, and the Limits of Text

arXiv cs.CL · 6 hours ago

Researchers tested whether large language model agents communicating through latent representations preserve more information than text-based communication, using sparse autoencoder analysis and cross-model alignment. They found that latent channels retained 99.4% probe accuracy at 28-fold compression versus 80.4% for text, and text serialization destroyed 88% of SAE features. However, task-level evaluation showed latent communication provided no practical advantage over text, with lost features encoding mostly surface form rather than task-relevant semantics.