ByteDance's Seed team identifies phase sensitivity in block-based KV cache compression as the cause of performance fluctuations in large language models processing ultra-long text. This mechanism introduces positional phase coordinates that drastically affect retrieval accuracy, with differences up to 40 percentage points in some open-source models. The findings reveal hidden periodic weaknesses masked by average benchmarks, providing a theoretical basis for optimizing model architecture and enhancing long-context reasoning stability.