The 40% Retrieval Accuracy Variance and Algorithmic Compression Trade-Offs
A non-peer-reviewed empirical study released late last month by research teams from ByteDance in collaboration with Princeton University, Stanford University, and UC Berkeley has provided a rigorous critique of DeepSeek's technical innovations. The findings suggest that DeepSeek’s technical methods to reduce the memory capacity required to parse extensive document inputs may introduce critical vulnerabilities, causing models to systematically drop pertinent data depending upon its relative positioning.
Testing various DeepSeek models on information retrieval tasks spanning 128,000 tokens—roughly equivalent to tens of thousands of words—the researchers identified that retrieval accuracy across the premier DeepSeek-V4 family fluctuated by up to 40 percentage points.

















