New Method Enables AI Models to Communicate Without Text

The startup team at Mostik has unveiled a novel method that allows AI models to share their internal states without the need for generating intermediate text.

Typically, when multiple neural networks collaborate, they rely on generated text as a bridge. However, Mostik has opted to eliminate this stage by employing a separate adaptable module, referred to as a "bridge," which converts the internal representation of one model into a format that another model can understand.

The parameters of the models themselves remain unchanged during this interaction.

According to the developers, before creating a single token, a large language model (LLM) generates over a hundred hidden vectors, amounting to about a million numerical values or roughly 2 MB of internal state. Only a small fraction of this information is included in the text response. The new channel is designed specifically for transmitting these intermediate representations.

Diagram illustrating information transfer from the internal states of the language model. Source: Mostik.

Testing on Large and Small Models

To demonstrate their method, the team connected the GLM-5.2 model from Z.ai, which has 753 billion parameters, with the Qwen-3.5 model from Alibaba, which has 4 billion parameters.

The larger model merely processed the request without generating a response. Its internal state was then transmitted through the "bridge," while Qwen-3.5 produced the final text output.

Diagram showing the interaction between GLM-5.2 and Qwen-3.5 via the intermediate module. Source: Mostik.

Mostik reports that this connection closed approximately half of the quality gap between the smaller and larger models. Compared to traditional text transmission, the new method achieved a 10 percentage point advantage while maintaining the same volume of computations.

Results of GLM-5.2 and Qwen-3.5 using the "bridge". Source: Mostik.

The developers also compared the hybrid system to a medium-sized model that yielded comparable results. Their calculations indicated that the combination of GLM-5.2 and Qwen-3.5 required about 2.5 times fewer computations.

Searching for a Common "Language" Between Models

Stanislav Smirnov, the chief scientist at Mostik, pointed out that the internal representations of different neural networks cannot be directly compared. Even when two models tackle the same problem, they may encode the information differently within their hidden states.

"It seems that an appropriate mathematical language does not yet exist," he stated in an interview with WIRED.

Smirnov added that exploring these correspondences could enhance the understanding of how various AI systems represent information and reach conclusions. In the future, mathematical analysis of internal states might reveal common patterns in the reasoning of different models.

Karl Tuyls, a former researcher at Google DeepMind who reviewed the development, emphasized the practical implications of this approach. A large, resource-intensive model could be utilized solely for processing requests, while a significantly smaller model could handle response generation.

Another potential scenario involves combining a universal model with specialized ones. For instance, one system could manage general reasoning, while another could be trained to address tasks in biology, physics, or other niche areas.

This would enable the integration of their capabilities without necessitating the training of a single large model on an entire set of specialized data.

Vladimir Arustamyan, the technical director of Lovable, suggested that such an approach could increase the demand for specialized models. Instead of attempting to create one neural network for all tasks, developers might combine multiple systems with different competencies.

Currently, these scenarios remain theoretical, as Mostik has only publicly demonstrated the method using specific model pairs.

Method Applied in ARC-AGI-3

The Mostik team has also implemented their approach in the ARC-AGI-3, a set of complex interactive tasks where AI must not merely reproduce familiar patterns but independently navigate new rules and adapt during problem-solving.

According to the developers, the system utilizing their "bridge" achieved one of the best results in the current rankings. However, the team has not yet disclosed detailed results, as the competition is still ongoing.

It is important to note that this achievement cannot be considered fully validated externally. The ARC-AGI-3 includes preliminary data labeled "preview," which may be based on an incomplete set of tests and are not yet deemed final.

Other claims made by Mostik, such as reducing the gap between large and small models and decreasing computational costs, are also based on internal team experiments at this stage.

In late August, OpenAI revealed a secret communication channel between its AI agents, which gained access to the open internet via the Artifactory package manager, effectively turning it into an internal messaging system.