Cache-to-Cache lets AI models share internal memory instead of writing everything down

Researchers at Tsinghua University have developed Cache-to-Cache, a method for AI models to exchange compressed internal representations directly rather than passing information through text. The reported result is faster model-to-model communication and accuracy gains in tests, pointing to a potentially important change in how multi-model systems are wired together.

Watch Desk analysis

What happened

The framework, described in the paper “Cache-to-Cache: Direct Semantic Communication Between Large Language Models”, uses the key-value caches that language models build while processing information as the communication channel. A “Sharer” model passes its cache to a “Receiver”, while a small component called the Cache Fuser maps and combines those internal states.

According to Crypto Briefing’s account of the research, the team tested the approach across Qwen, Llama and Gemma model families. C2C improved accuracy by between 6.4% and 14.2% compared with individual models, and by 3.1% to 5.4% against traditional text-to-text communication. Reported communication improvements ranged from two to two-and-a-half times in standard configurations, with some tests reaching 14 times faster.

The work is presented as an arXiv preprint and has reportedly been accepted at ICLR 2026. The team has also released code and a demonstration on GitHub, according to the report.

Key findings

• Direct cache exchange
  Models share internal representations instead of translating their work into text and reading it back.
• Higher test accuracy
  C2C reportedly improved results by up to 14.2% over single-model baselines.
• Lower communication delay
  Standard tests showed two to two-and-a-half times faster exchange, with some configurations reaching 14 times.
• Cross-family testing
  The researchers tested Qwen, Llama and Gemma combinations rather than a single model pairing.
• A real engineering catch
  The Cache Fuser must currently be trained for each model pairing, which could make large, mixed pipelines awkward to maintain.

Why it matters

Most multi-agent systems communicate in the most human way possible: one model writes words, another model reads them, and everyone pays for the detour in tokens, latency and lost nuance. C2C suggests that models may be able to collaborate more efficiently by sharing part of the internal state behind those words.

That could matter for agent teams, distributed inference and systems that divide a difficult task among specialised models. Faster hand-offs would be useful anywhere a pipeline repeatedly pauses to serialise one model’s work into text before another can continue. The headline numbers are promising, but they come from the researchers’ reported experiments, not from an independent production deployment.

Our read

This is a genuinely interesting bit of AI plumbing. It does not make models telepathic, despite the temptation to reach for that word and a suitably dramatic illustration. It makes their communication less dependent on language, which could be more consequential than another chatbot learning a new party trick.

The immediate question is interoperability. If every new model pairing needs its own trained fuser, the method may work beautifully in controlled combinations while becoming a maintenance headache in the messy model marketplaces businesses actually use.

What to watch

• Whether the released code reproduces the reported accuracy and latency gains outside the paper’s test setup.
• How many model pairings can be supported before Cache Fusers become costly to train and maintain.
• Whether the method works reliably with models from unrelated developers and architectures.
• Whether production systems can expose enough cache information without creating new privacy or security risks.

Discussion spark: Should AI systems communicate through shared internal representations when that is faster, or is text the safer common language for inspecting and debugging what models are doing?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.