Cache-to-Cache: LLM Agents Share Context Without Text
Researchers introduced Cache-to-Cache, a method where language models exchange KV cache directly instead of converting context back to text. Accuracy improved by 6.4–14.2%, speed by roughly 2.5x, and the paper was accepted at ICLR 2026.
Researchers proposed Cache-to-Cache, a method where language models exchange their internal KV cache directly instead of going through text. Accuracy on benchmark tasks improved by 6.4–14.2% compared to individual models. Speed increased by roughly 2.5x. The paper was accepted at ICLR 2026.
Right now, passing context from one agent to another means converting it back to text, and that costs tokens. Cache-to-Cache skips that step by transferring the internal state directly. The method is still in the research phase and is not available in any production tools, but I think if it reaches a product, agent pipelines will become both cheaper and faster.
Source: arxiv.org
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides
Author
Evgenii Arsentev
PhD · Chief Executive Officer, digital health
Articles · Latest articles