Abstract
Agentic workflows powered by large language models are increasingly central to professional work, but natural-language communication between agents raises token usage, computational cost, and context-window pressure. This paper explores latent communication as a compact alternative to text-based exchange.
We implement a lightweight language model, TinyLM, and develop a sender-receiver architecture in which agents transmit only the final hidden state of the preceding sequence. On a synthetic Sally-Anne Theory of Mind dataset, the receiver achieves high query-answering accuracy and substantially outperforms zero-message and shuffled-message baselines.
The results show that a single hidden-state vector can preserve task-relevant information and serve as a viable communication medium. The paper also identifies limitations, including potential overfitting and the simplicity of the experimental setup, and outlines future work on scalability, more complex reasoning tasks, and information stability across multi-agent chains.