What's this about?
In multi-agent systems, LLM agents normally communicate in text. This paper argues text is a discrete bottleneck: converting the sender's continuous internal state into words discards information. StateBridge is a training-free method for aligning hidden states so agents can communicate in latent space instead.
What exactly is a "hidden state"?
As a model processes text it builds up a rich internal numerical representation — thousands of continuous values encoding meaning, uncertainty, and context far more finely than any sentence it eventually writes. When it produces output, all of that is collapsed into a sequence of discrete word choices. If a second agent then reads those words, it must reconstruct an internal state from scratch, and much of the original nuance is gone. Passing the internal representation directly is like handing someone your mental model instead of your best attempt at describing it.
Why is this worth noticing?
- Multi-agent systems are increasingly standard, and text is currently their only interface — so a lossless channel between agents could improve them without any change to the models.
- It's training-free, meaning it works on existing models rather than requiring a retraining run, which is what determines whether a technique gets adopted.
- The trade-off deserves stating: text between agents is human-readable, and that readability is a large part of how anyone debugs or audits a multi-agent system. Latent communication buys fidelity at the cost of legibility.