LmCast :: Stay tuned in

Cache-to-Cache: Direct Semantic Communication Between Large Language Models

Recorded: Sept. 18, 2026, 8 p.m.

Original Summarized

[2510.03215] Cache-to-Cache: Direct Semantic Communication Between Large Language Models

Skip to main content

Search

Submit
Donate

Log in

Search arXiv

Press Enter to search · Advanced search

Computer Science > Computation and Language

arXiv:2510.03215 (cs)

[Submitted on 3 Oct 2025 (v1), last revised 2 Mar 2026 (this version, v2)]
Title:Cache-to-Cache: Direct Semantic Communication Between Large Language Models
Authors:Tianyu Fu, Zihan Min, Hanling Zhang, Jichao Yan, Guohao Dai, Wanli Ouyang, Yu Wang View a PDF of the paper titled Cache-to-Cache: Direct Semantic Communication Between Large Language Models, by Tianyu Fu and 6 other authors
View PDF
HTML (experimental)

Abstract:Multi-LLM systems harness the complementary strengths of diverse Large Language Models, achieving performance and efficiency gains that are not attainable by a single model. In existing designs, LLMs communicate through text, forcing internal representations to be transformed into output token sequences. This process both loses rich semantic information and incurs token-by-token generation latency. Motivated by these limitations, we ask: Can LLMs communicate beyond text? Oracle experiments show that enriching the KV-Cache semantics can improve response quality without increasing cache size, supporting KV-Cache as an effective medium for inter-model communication. Thus, we propose Cache-to-Cache (C2C), a new paradigm for direct semantic communication between LLMs. C2C uses a neural network to project and fuse the source model's KV-cache with that of the target model to enable direct semantic transfer. A learnable gating mechanism selects the target layers that benefit from cache communication. Compared with text communication, C2C utilizes the deep, specialized semantics from both models, while avoiding explicit intermediate text generation. Experiments show that C2C achieves 6.4-14.2% higher average accuracy than individual models. It further outperforms the text communication paradigm by approximately 3.1-5.4%, while delivering an average 2.5x speedup in latency. Our code is available at this https URL.


Comments:
Published in ICLR'26

Subjects:

Computation and Language (cs.CL); Machine Learning (cs.LG)

MSC classes:
68T07, 68T50

ACM classes:
I.2.7

Cite as:
arXiv:2510.03215 [cs.CL]

 
(or
arXiv:2510.03215v2 [cs.CL] for this version)

 
https://doi.org/10.48550/arXiv.2510.03215

Focus to learn more

arXiv-issued DOI via DataCite

Submission history From: Tianyu Fu [view email] [v1]
Fri, 3 Oct 2025 17:52:32 UTC (484 KB)
[v2]
Mon, 2 Mar 2026 19:24:02 UTC (546 KB)

Full-text links:
Access Paper:

View a PDF of the paper titled Cache-to-Cache: Direct Semantic Communication Between Large Language Models, by Tianyu Fu and 6 other authorsView PDFHTML (experimental)TeX Source

view license


Current browse context:
cs.CL

< prev

  |  
next >

new
|
recent
| 2025-10

Change to browse by:

cs
cs.LG

References & Citations

NASA ADSGoogle Scholar
Semantic Scholar

export BibTeX citation
Loading...

BibTeX formatted citation
×

loading...

Data provided by:

Bookmark

Bibliographic Tools

Bibliographic and Citation Tools

Bibliographic Explorer Toggle

Bibliographic Explorer (What is the Explorer?)

Connected Papers Toggle

Connected Papers (What is Connected Papers?)

Litmaps Toggle

Litmaps (What is Litmaps?)

scite.ai Toggle

scite Smart Citations (What are Smart Citations?)

Code, Data, Media

Code, Data and Media Associated with this Article

alphaXiv Toggle

alphaXiv (What is alphaXiv?)

Links to Code Toggle

CatalyzeX Code Finder for Papers (What is CatalyzeX?)

DagsHub Toggle

DagsHub (What is DagsHub?)

GotitPub Toggle

Gotit.pub (What is GotitPub?)

Huggingface Toggle

Hugging Face (What is Huggingface?)

ScienceCast Toggle

ScienceCast (What is ScienceCast?)

Demos

Demos

Replicate Toggle

Replicate (What is Replicate?)

Spaces Toggle

Hugging Face Spaces (What is Spaces?)

Spaces Toggle

TXYZ.AI (What is TXYZ.AI?)

Related Papers

Recommenders and Search Tools

Link to Influence Flower

Influence Flower (What are Influence Flowers?)

Core recommender toggle

CORE Recommender (What is CORE?)

Author
Venue
Institution
Topic

About arXivLabs

arXivLabs: experimental projects with community collaborators
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

Which authors of this paper are endorsers? |
Disable MathJax (What is MathJax?)

We gratefully acknowledge support from
our major funders,
member institutions, ,
and all contributors.

About
·
Help
·
Contact
·
Subscribe
·
Copyright
·
Privacy
·
Accessibility
·
Operational Status (opens in new tab)

Major funding support from

Multi-Large Language Model systems leverage the complementary strengths of diverse models to achieve performance and efficiency gains unattainable by single models. Current methods for communication between Large Language Models typically rely on text-based exchange, which necessitates transforming internal representations into output token sequences. This process results in the loss of rich semantic information and introduces generation latency due to token-by-token processing. Researchers investigate whether LLMs can communicate beyond the limitations of text. Oracle experiments indicate that enhancing the semantics within the Key-Value (KV) Cache can improve response quality without increasing the cache size, thereby positioning the KV-Cache as an effective medium for inter-model communication. Building upon this, the authors propose Cache-to-Cache (C2C), a novel paradigm designed for direct semantic communication between LLMs.

The core of the C2C approach involves employing a neural network to project and fuse the KV-caches from the source model and the target model, facilitating direct semantic transfer. To optimize this communication, the method incorporates a learnable gating mechanism that selectively determines which layers in the target model benefit most from this cache communication. This mechanism allows the system to selectively focus the semantic transfer where it is most beneficial. By leveraging the deep and specialized semantics embedded in both models, C2C avoids the need for explicit intermediate text generation, which is characteristic of text communication methods.

Empirical evaluations demonstrate the superiority of the C2C method. The experiments show that C2C achieves an average accuracy improvement ranging from 6.4 to 14.2% compared to individual models. Furthermore, C2C surpasses the text communication paradigm by approximately 3.1 to 5.4%, while simultaneously delivering an average speedup in latency of 2.5 times. This indicates that direct semantic communication via cache fusion is not only more effective in capturing shared knowledge but also significantly more efficient in terms of computational time.