C²KV: Compressed and Composable KV Cache Reuse for Efficient LLM Inference¶
Small NumPy core diagnostic, not paper-scale inference.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Shanghai Jiao Tong University; first author Chuheng Du. |
| First public release | 2026-07-20 (arXiv v1) |
| Author code | s7a9/C2KV. |
| Local adapter or method key | c2kv |
| Local reproduction code | src/auto_research/reproductions/c2kv/ |
Paper method and evidence¶
Learn position-independent, composable KV representations with compression tokens and structured attention, jointly training compression and reuse.
flowchart LR
A[Compression tokens] --> B[Position-independent KV]
B[Position-independent KV] --> C[Compose reusable segments]Original paper figure¶
Figure shows composable cache reuse. From the original paper; copyright remains with the authors.
Core equation¶
\[
Z=\operatorname{Extractor}(KV),\quad \hat y=F([Z_1;\ldots;Z_m])
\]
Reported results¶
The paper reports up to 17× inference speedup across long-context benchmarks while retaining generation quality.
Local reproduction¶
Same 64 examples and seed: recent-window baseline 0.0781, C²KV 0.8438, +76.56 percentage points. Metrics, batch index. No paper-scale checkpoint, custom CUDA kernel or full benchmark.