Byte Seed team found DeepSeek "pumping" reason, long context probability can drift字节 Seed 团队发现 DeepSeek“抽风”原因，长上下文可能性能漂移

IT House October 9th news, byte Seed team submitted a paper at the end of September this year on the preprint platform arXiv, talking about the phase sensitivity caused by block KV cache compression, pointing directly to the DeepSeek "pumping" reason. The study evaluated basic and post-trained DeepSeek-V4-Flash and DeepSeek-V4-Pro, as well as post-trained DeepSeek-V4.1-Flash. The model compresses consecutively marked windows into fewer cache entries by blocking KV cache compression to a fixed stride, reducing the memory and attention costs of long contextual reasoning. However, this compression also introduces a new position coordinate: the phase of the token, or its position relative to the boundary of the compressed window. The team found a systemic asymmetry in models using this compression: the same information is easy to retrieve at one stage, but difficult to retrieve at another. The team refers to this periodic change in retrieval performance as phase sensitivity. In large open weight models employing such compression, long context retrieval accuracy can vary by up to 40 percentage points between stages, revealing periodic weaknesses that average benchmark scores may mask. IT Home with thesis link: https://arxiv.org/abs/2609.36322