Study examines reuse and negative transfer in latent communication between model cells

A new arXiv paper looks at how groups of language models that share a common base communicate through compact latent packets rather than plain text. The authors find that limiting what each cell can see encourages reusable, value-indexed interfaces, while a single globally visible model instead developed a code entangled with specific episodes. The work highlights both the transfer benefits and the negative transfer risks when latent messages are reused across different settings.