The Seed team at ByteDance published a paper on arXiv in late September regarding the performance of DeepSeek models. The study specifically analyzed the DeepSeek-V4-Flash and DeepSeek-V4-Pro versions, as well as the post-trained DeepSeek-V4.1-Flash.
Researchers investigated the impact of chunked KV cache compression, a technique used to lower memory and attention costs during long-context inference. This method compresses continuous token windows into fewer cache entries using a fixed stride.
The study identified a phenomenon called phase sensitivity, which introduces a new positional coordinate based on a token's location relative to compression window boundaries. This creates a systematic asymmetry where information retrieval performance fluctuates.
The team observed that retrieval accuracy in models using this compression can vary by up to 40 percentage points across different phases. This periodic weakness is often hidden by average benchmark scores, highlighting potential reliability issues in long-context tasks.
