What changes when two AI models communicate through internal signals instead of words?
An independent researcher trained two small models from different families to exchange hidden activation vectors rather than text, bridged by LoRA adapters. The code and weights—the learned model parameters—are open on Hugging Face and GitHub.
Decoding the idea
Activation vectors are numerical representations inside a model. In this experiment, those representations carry the communication between models.
LoRA adapters are small, trainable components that connect the two models in this setup. The detail worth noticing: the models come from different families, yet were trained to communicate through these internal representations.
A practical next step
For practitioners, the released code and weights offer a starting point for examining how the connection works. A useful evaluation would compare task results with text-based communication and examine how to inspect the signals exchanged.
Treat this as an experiment to evaluate against a specific workflow before adopting the approach.