Cache-to-Cache
A research approach for cross-model communication in multi-agent systems that transmits KV Cache (key-value cache) directly between models instead of relying on intermediate text layers, enabling higher-fidelity and lower-latency semantic transfer. The core challenge lies in resolving the alignment of representation spaces across heterogeneous models.
Timeline (last 90 days)
Cache-to-Cache (C2C) proposes that models bypass the text intermediate layer and conduct semantic transfer directly at the KV Cache level.
The Cache-to-Cache work received 42 points on Hacker News.
C2C establishes a mapping mechanism that converts and aligns the source model's KV Cache into a form directly usable by the target model, allowing the target model to take over the semantic understanding already completed by the source model.
The representation spaces of different models are not naturally aligned—their architectures, dimensions, and training data may all differ—which is the core technical challenge C2C needs to solve.
Cache-level communication sacrifices the readability and interpretability of text; humans cannot directly understand the content being transferred, which increases the difficulty of debugging and security auditing.
C2C offers two benefits: bypassing the lossy text channel to preserve more complete semantics, and eliminating the need to redo prefill, thereby saving redundant computation and reducing latency.
Whether cross-model representation alignment in C2C is robust and whether it applies to highly heterogeneous models with large architectural differences still requires further validation.
All Facts (6)
Cache-to-Cache (C2C) proposes that models bypass the text intermediate layer and conduct semantic transfer directly at the KV Cache level.
50%UnverifiedC2C establishes a mapping mechanism that converts and aligns the source model's KV Cache into a form directly usable by the target model, allowing the target model to take over the semantic understanding already completed by the source model.
50%UnverifiedThe representation spaces of different models are not naturally aligned—their architectures, dimensions, and training data may all differ—which is the core technical challenge C2C needs to solve.
50%UnverifiedCache-level communication sacrifices the readability and interpretability of text; humans cannot directly understand the content being transferred, which increases the difficulty of debugging and security auditing.
50%UnverifiedC2C offers two benefits: bypassing the lossy text channel to preserve more complete semantics, and eliminating the need to redo prefill, thereby saving redundant computation and reducing latency.
50%UnverifiedWhether cross-model representation alignment in C2C is robust and whether it applies to highly heterogeneous models with large architectural differences still requires further validation.
50%