Multi-agent systems composed of language models can tackle tasks by sharing complementary information. Traditionally, these agents communicate via natural language text, but recent work has explored exchanging internal model states—continuous vectors—to convey more information.
The new paper, "Share the Symbol, Not the Vector," proposes an alternative: communicate facts by sharing discrete symbols rather than continuous vectors. The method is training-free, meaning it does not require additional fine-tuning of the models to enable this form of communication.
The abstract is brief and does not include experimental details or comparisons. Still, the central idea—favoring symbolic exchange over vector-based messaging—offers a distinct direction for research into inter-model communication.