The authors formulate zero-prefill cross-model context handoff and conduct diagnostic analyses of KV reuse, showing that similar reconstruction errors can lead to markedly different receiver behavior.
Functional supervision improves cross-model KV cache transport over reconstruction alone
CacheLink learns a coupled K/V correction through the frozen receiver, reducing handoff latency by up to 7.27x at 32K context.
Top University
Haomiao Qiu · Yining Tang · Zijia Song · Ruizi Han · Ziyue Qiao · Miao Zhang
Harbin Institute of Technology · Shenzhen Great Bay University · Zhejiang University
Research Digest··2 min read
The authors study zero-prefill cross-model context handoff, where a sender's KV cache is translated for direct receiver use.
Why this paper
From Zhejiang University and 2 others
In one line
CacheLink achieves zero-prefill cross-model KV handoff by optimizing receiver computation rather than tensor fidelity, reducing latency up to 7.27x.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§