Researchers Release C2C to Let AI Models Share Internal Memory Instead of Text

Researchers from Tsinghua University and other Chinese institutions have released Cache-to-Cache, or C2C, a method that lets language models pass information through internal memory states rather than generated text. In the reported tests, it beat text handoffs on accuracy and latency—but only in controlled benchmarks with access to the models’ inner workings.

The work targets a recurring inefficiency in systems that route work between models or assign them different roles. Normally, a first model turns its understanding into a sequence of tokens; the next model then reads those tokens and reconstructs their meaning. That relay takes time and can discard details embedded in the first model’s internal representation.

A relay without the message

C2C uses the key-value cache, or KV cache: the stored internal states a language model creates while processing a prompt and reuses while generating an answer. Both participating models process the shared context. A trained cache-fusion module then maps the Sharer model’s cache into the Receiver model’s representation space and combines it with the Receiver’s cache. A learned gate chooses which Receiver layers get that added information.

The paper’s results are strongest as a proof of the communication idea, not as a deployment verdict. With Qwen3-0.6B fixed as the Receiver, C2C improved average accuracy by roughly 3.1 to 5.4 percentage points over text-to-text communication. The official ICLR abstract reports an average 2.5× latency speedup, while the more detailed tests reported substantially different gains across model pairs.

A bridge, not a drop-in feature

C2C was evaluated on MMLU-Redux, OpenBookQA, ARC-Challenge and C-Eval, including configurations that paired different model families or specialized models. The researchers froze both language models and trained only the fusion module, avoiding a full fine-tune of either participant. But a bridge still has to account for incompatible tokenizers, layer counts and representation sizes.

What teams would need

• An inference stack that can read, transform and replace internal KV-cache states.
• A trained cache-fusion bridge for the models they want to connect.
• Validation beyond the reported academic benchmarks before relying on it for agent workloads.

That leaves it most relevant today to teams that operate their own inference infrastructure, rather than applications that connect closed model APIs. The researchers have released the code and configurations under an Apache 2.0 license, plus pretrained cache-fuser checkpoints. Whether the efficiency gains survive real, messy agent workflows remains the central unanswered question.