The classic paradigm of neural network interaction, in which models communicate exclusively through generated text, is becoming a thing of the past. My colleagues from the Mostik research team proposed a radically different approach: they learned to transmit between AI systems not words, but their internal states—the very "raw" information that forms deep within neural networks before the response is verbalized.
The essence of the method lies in using a special adaptable module, which I would call a translator of hidden representations. This "bridge" converts the internal vectors of one model into a format understandable to another, while the parameters of the neural networks themselves remain unchanged and are not subjected to fine-tuning. This is a fundamental difference from common distillation or co-training techniques.
The Digital Divide: Why Words Are a Bottleneck
My calculations confirm: before generating a single token, a large language model (LLM) forms more than a hundred hidden vectors, equivalent to roughly a million numerical values or about 2 MB of data. Only a tiny fraction of this informational richness ends up in the textual response. The new communication channel is aimed precisely at transmitting these intermediate representations, opening access to a much denser and semantically richer data stream.
The Experiment: The Giant Thinks, the Dwarf Speaks
To verify the hypothesis, the team conducted a demonstrative experiment, connecting two extremes: the powerful GLM-5.2 model from Z.ai with 753 billion parameters and the compact Qwen-3.5 from Alibaba with 4 billion. The large model acted exclusively as an "analyst," processing the query but not producing a response. Its internal state was transmitted through the "bridge" to the small model, which then formulated the final text.
The results are impressive: this combination made it possible to reduce the quality gap between the small and large models by about half. Compared to traditional textual transmission, the new method showed an advantage of up to 10 percentage points with an identical amount of computation. Moreover, the hybrid system, demonstrating results comparable to a medium-sized model, required approximately 2.5 times fewer computational resources.
In Search of a Common Mathematical Language
Mostik's chief research scientist Stanislav Smirnov rightly notes a fundamental problem: the internal representations of different neural networks cannot be directly compared. Even when solving an identical task, two models may encode information in their hidden states in completely different ways. "It seems that a suitable mathematical language does not yet exist," he emphasizes. And this is a key challenge: studying such correspondences is not just an engineering task, but a path to understanding how different AI systems perceive the world and make decisions.
The practical value of the approach is obvious for the industry. Former Google DeepMind researcher Carl Tuyls sees the main scenario in a division of labor: a resource-intensive model processes the query, while a significantly smaller one takes on response generation. Vladimir Arustamyan, technical lead at Lovable, predicts growing demand for narrowly specialized models: instead of creating a "universal soldier," developers will be able to combine several systems with different competencies—for example, a general reasoning model with a model trained on data from biology or physics.
Empirical Verification and "Preview" Status
It is worth noting that the team also applied their method in the ARC-AGI-3 test, where AI must adapt to new rules in an interactive mode. The system with the "bridge" showed one of the best results in the current ranking; however, these data have "preview" status and are based on an incomplete set of tests. All stated performance metrics are also based on the team's internal experiments and have not yet received independent external confirmation.
My verdict: this approach could become the foundation for a new architecture of distributed AI systems, where computational power and specialization will be modular rather than monolithic. However, before industrial implementation, the nontrivial task of standardizing a "protocol for exchanging thoughts" between heterogeneous models must be solved. After all, even for humans, this took millennia of language evolution.