Neural Twins Quill and Ember Talk Through Activations, Not Words

In multi-agent systems, models still talk in a fairly primitive way: one writes its thoughts out as text, and the other reads them back in. A lot of information is lost the moment ideas are squeezed into tokens. An experimental project just open-sourced by independent developer Sauers tries to skip that step and let two models talk through something closer to brainwaves.

The project, neural-pairing, centers on a pair of "neural twins": Quill, built on Qwen3-4B, and Ember, built on Arcee's AFM-4.5B. They don't exchange text. At eight layer checkpoints, each model attends to its partner's past hidden states and injects gated, normalized activations into its own residual stream. In effect, each model can read the other's thinking while it thinks.

A "neuralese" trained through games

To make the channel carry useful information, the pair were trained on communication games. In one, a model gets a hidden word and has to pass it to its partner through activations alone. After training, they can also take turns writing a single interleaved piece of text. Sauers calls this activation-level, human-unreadable exchange "neuralese."

The side effects are the interesting part. Sauers reports cross-model concept steering, where activating a concept in one model pulls the other's output toward it, and something like emotion transfer. The model card is careful here: the "feelings" are tendencies measured in generated text, not evidence of experience.

A small experiment with big questions

As engineering, this is a toy-scale experiment with two 4B-class models, far from practical use. But it touches a frontier problem. If models can share representations directly instead of going through language, multi-agent collaboration could get far more bandwidth with lower latency and token cost, which is exactly where agent systems hurt most today.

The other side is interpretability and safety. One big reason we can still supervise agents working together is that their conversation is readable by humans. Once communication moves into neuralese, auditing gets much harder. Safety researchers have long worried about models developing private channels people can't read, and Quill and Ember offer a public specimen to take apart.

It's also notable that the two models come from different architectures and different labs. That suggests the "translation" between activation spaces can be learned, and the two sides don't need a common origin.

Weights, code and training recipes are on Hugging Face and GitHub. For anyone working on multi-agent communication, representation alignment or AI safety, it's a starting point you can actually poke at, and a reminder worth taking seriously: conversations between models may not always happen in a language we can read.