DeepSeek's memory-saving technique may reduce accuracy, study finds
AIA non-peer-reviewed paper led by ByteDance researchers with Princeton, Stanford and UC Berkeley affiliates says DeepSeek's memory-reduction method for long inputs may make models miss information in certain document positions. The paper says the effect appears in DeepSeek-V4, V4.1 and all compressed models the researchers pretrained.
Source details
South China Morning Post · Tech · Minxiao Chang · scmp.comPublished · added here
Original title: DeepSeek’s cost-saving AI may lead to inaccuracy in model performance: research