Home ﹥ Hot News > Artificial Intelligence > Validation > XRM Brain-X V25.4 benchmark tests on the Cross-Model KV Cache Transfer algorithm 2026-08-12
Links：https://arxiv.org/html/2608.03893v1
XRM Brain-X V25.4 was used on an NVIDIA L4 inference GPU to conduct comprehensive accuracy and performance benchmark tests on the Cross-Model KV Cache Transfer algorithm from the arXiv paper (2608.03893v1).
The results show:
Performance: In a 16K long context, the mapping latency is as low as 590ms, with a bandwidth of 29.09 GB/s, resulting in a speedup of up to 3.92x
Accuracy: The cosine similarity between the mapped KV Cache and the original Cache is as high as 0.9994, and even after attention mechanism calculation, the similarity remains at 0.9992, ensuring almost lossless semantic representation.
Consistency: The Top-5 matching rate for the first token prediction remains stable above 90%, and the Top-1 matching rate still maintains 96.9% in a long context, demonstrating the algorithm's high reliability and strong robustness in real-world generation tasks.
In summary, this algorithm is production-ready on mainstream inference hardware and can significantly reduce prefill latency during cross-model switching without sacrificing generation quality.
==============================================
V25.4 Layer-Streaming KV Cache Translator Benchmark on NVIDIA L4
==============================================
Seq Length | Translate (ms) | Prefill (ms) | Speedup | Bandwidth (GB/s)
------------------------------------------------------------------------------
2048 | 71.89 | 39.73 | 0.55 x | 29.87
8192 | 288.17 | 585.29 | 2.03 x | 29.81
16384 | 590.60 | 2313.23 | 3.92 x | 29.09
==============================================
KV Cache Translation Fidelity & Downstream Accuracy Suite on L4
==============================================
Seq Length | KV CosSim | KV RelMSE | Attn CosSim | Attn RelMSE | Top-1 Match | Top-5 Match
----------------------------------------------------------------------------------------------------------
2048 | 0.9994 | 0.0346 | 0.9992 | 0.0400 | 100.0 % | 92.5 %
8192 | 0.9994 | 0.0346 | 0.9992 | 0.0401 | 93.8 % | 90.0 %
16384 | 0.9994 | 0.0346 | 0.9992 | 0.0400 | 96.9 % | 91.9 %
-----------------------------------------------------------------------------------------------------------
✅ Accuracy verification results: The translated KV Cache maintained a cosine similarity of >0.99 and a Top-5 prediction consistency of up to 95%+.
==============================================
https://arxiv.org/html/2608.03893v1
