Latent Communication Interface (Lci): A Framework For Direct Representation-Level Communication Between Specialized Transformer Models

Abstract

Modern AI systems depend on ensembles of specialized transformer models that are fine-tuned for specific tasks including reasoning, instruction execution, coding, mathematics, and multimodal perception. It is important that one model converts its internal state into tokens and another model converts them back into distributed representation, albeit the models have the same origin, i.e. they have been pre-trained in the same way. The method is time consuming, complex and also introduces unwanted delays in the communication between the models. In this study, we will specify the situations in which latent communication can be considered as an alternate way to token-based communication between models and characterize the field conditions, statements and constraints of this proof-of-concept research. The main contribution of this paper is a theoretical framework and an exhaustive research design for a same-family LLM pair (Section 5). Further, we present a small-scale empirical pilot - deployed and evaluated end-to-end using toy transformer models with a shared pretrained basis - that gives preliminary, but inconclusive data on the operability of the mechanism. Results are provided candidly at the scale they were attained and should not be taken as indication of production-LLM viability.