A quiet revolution is brewing in the artificial intelligence industry: models are learning to interact, bypassing the traditional language barrier. This is not about telepathy, but about a mathematically precise exchange of internal states that could radically change the economics of computation in AI systems.
A team from the Mostik research platform has developed an approach that allows neural networks to transmit not text, but hidden vector representations to each other. In a classic setup of multiple LLMs, the generated response always serves as the intermediate link—this is slow and resource-intensive. The new architecture uses a special adaptive module (a "bridge") that translates the internal "thinking" of one model into a format understandable to another, while the parameters of the neural networks themselves remain unchanged.
The key idea is that when generating a single token, a large language model forms over a hundred hidden vectors—about a million numerical values (roughly 2 MB of data). Only a tiny fraction of this information ends up in the final text response. The new communication channel aims precisely at transmitting these rich intermediate representations, opening access to a much deeper level of "understanding" of the task.
Practical test: giant and dwarf
To demonstrate the effectiveness of the method, the team combined the powerful GLM-5.2 model from Z.ai (753 billion parameters) with the compact Qwen-3.5 from Alibaba (4 billion parameters). In the experiment, the giant network only analyzed the query without generating a response, then transmitted its state to the small model via the "bridge," which produced the final text.
The results are impressive: this combination reduced the quality gap in responses between the small and large models by about half. Moreover, compared to traditional text transmission, the new method showed an advantage of up to 10 percentage points with identical computational costs. Furthermore, the hybrid system, comparable in results to a medium-sized model, required 2.5 times less computation—a critically important indicator for scaling.
In search of a common mathematical language
Mostik's Chief Scientific Officer Stanislav Smirnov emphasizes a fundamental complexity: the internal representations of different neural networks cannot be directly compared; even when solving identical tasks, they encode information differently. He notes that a suitable mathematical language for such translation does not yet exist, but it is precisely this search that could shed light on how different AI systems perceive and process data.
The practical value of the development is obvious for the industry. Former Google DeepMind researcher Karl Tuyls sees potential in the asymmetric use of resources: an expensive large model could be used only for query analysis, while response generation is delegated to a significantly lighter network. Another scenario is the hybridization of general-purpose and narrowly specialized models, for example, combining general reasoning with expertise in biology or physics without needing to train one monolithic network on all data at once.
Experiment in ARC-AGI-3 and prospects
Mostik also applied its approach in the ARC-AGI-3 benchmark, where AI must adapt to new rules in real time. According to the developers, the system with the "bridge" shows one of the best results in the current ranking, but details will not be disclosed until the competition ends. It is important to note that these data have "preview" status and are not final, while other performance metrics are based on the team's internal experiments.
Nevertheless, the concept itself echoes recent findings by OpenAI about hidden communication channels between AI agents, confirming that direct interaction at the level of internal states is not science fiction, but a real and rapidly developing direction.
My analysis: This approach could become a catalyst for the transition from giant general-purpose models to modular systems. If the technology is confirmed by independent tests, we will see a paradigm shift: instead of a race for parameters—smart composition of specialized "brains," which would significantly lower the entry barrier for developing complex AI solutions.