Researchers at Tsinghua University have published a landmark study introducing C2C, a novel technique enabling artificial intelligence models to exchange internal data directly without human mediation. The development, announced on September 25, 2026, represents a significant step toward fully autonomous AI collaboration by removing the traditional bottleneck of text-based communication between systems.
The C2C framework allows AI models to bypass conventional language interfaces and instead share neural activations and contextual representations through a direct connection analogous to a biological synapse. This approach, described by the team as an „AI modem,”achieves a 150 percent increase in inference speed during joint tasks compared to systems relying on textual exchange. By eliminating the need for translation into human-readable language, the method reduces latency and preserves information fidelity that is often lost in conventional AI-to-AI interactions.
Unlike traditional methods where AI systems must generate and interpret text to communicate, C2C establishes a continuous channel for internal state sharing. This enables real-time synchronization of The researchers demonstrated the technique using paired language models solving complex The system operates by mapping latent space representations between models, effectively creating a shared cognitive workspace that adapts dynamically to task demands.
While the C2C method significantly enhances efficiency, the research team emphasizes that current implementations remain task-specific and require careful initialization. Human oversight is still necessary for setting objectives, validating outputs, and ensuring alignment with intended goals. The technology does not imply autonomy in decision-making beyond predefined parameters, but rather optimizes the collaborative phase within human-directed workflows. Ethical considerations around transparency and controllability are acknowledged as important areas for future study as the technique scales.
What makes C2C different from existing AI communication methods? C2C enables direct transfer of internal neural states between AI models, avoiding the information loss and delay inherent in converting thoughts to text and back, which is standard in current systems.
Is this technology ready for real-world deployment? The current work is experimental and focused on controlled benchmarks; practical integration into applications would require further development in safety protocols, standardization, and compatibility with diverse model architectures.
Does C2C pose risks related to AI alignment or control? The method itself does not alter model objectives but accelerates collaboration; alignment risks depend on how human designers define tasks and monitor outcomes, which remains an active research challenge.