Khaberni - Artificial intelligence models face a critical constraint known as the bottleneck for long texts, which limits their ability to process lengthy documents.
A group of researchers from China and Japan challenged a method unveiled several months ago by the emerging Chinese artificial intelligence company "DeepSeek," which was designed to improve AI's handling of long texts, marking a rare case where the company's research is publicly questioned.
According to researchers from Tohoku University in Japan and the Chinese Academy of Sciences, the "DeepSeek-OCR" methodology (Optical Character Recognition), designed to compress texts using visual representations—which was supposed to revolutionize how AI models handle long texts—is flawed due to inconsistent performance, according to a report by the South China Morning Post, reviewed by "Al Arabiya Business".
In their study, the research team found that "DeepSeek's" method relies heavily on prior linguistic knowledge – meaning AI models' tendency to rely on patterns learned from massive amounts of text – rather than the visual understanding that the company claims to achieve, making the performance metrics announced by the Chinese company "misleading".
The researchers pointed out that artificial intelligence models face a critical constraint known as the "long context bottleneck," which restricts their ability to process long documents or prolonged conversations.
Companies and research institutes around the world have sought to achieve improvements in this area, due to its significant impact on enhancing the performance of AI systems.
It was stated that the "DeepSeek-OCR" technology, published in October, is capable of handling large and complex documents using visual perception as a means of compression.
The company said at the time: "Visual context compression can achieve a significant reduction in the number of symbols - from seven to twenty times... presenting a promising direction" for dealing with long context challenges in artificial intelligence.
However, new research, in a series of carefully designed experiments, found that the accuracy of "DeepSeek-OCR" responses to visual questions dropped to about 20% when provided with additional text to influence its inference, compared to an accuracy of over 90% for standard AI models.
The researchers said this gap "ultimately raises questions about whether the current methods of visual compression represent a viable path to solving long-context constraints in AI models," and indicated that alternative strategies might be necessary.
Conversely, some computer scientists described the "DeepSeek-OCR" algorithm as a double-edged sword more than an inherent flaw, as there is no magic solution for all scenarios.
Li Boji, who earned his PhD in computer science from the University of Science and Technology of China and currently runs his startup in the field of artificial intelligence in Beijing, said that for manuscripts that are difficult to recognize, relying on acquired knowledge may help AI understand the text, but it could be a drawback for clearly printed materials.
