A technology developed by the Chinese AI startup “DeepSeek” has sparked a heated scientific debate after a new research study cast doubt on its claimed ability to improve how AI models handle lengthy texts, in a rare incident where the company’s research is facing public criticism.
The study, supervised by researchers from Tohoku University in Japan and the Chinese Academy of Sciences, focused on “DeepSeek-OCR” technology, a mechanism that relies on visually representing texts to reduce their size, which is supposed to help AI models overcome the “long context bottleneck.”
What is “DeepSeek-OCR” technology?
“DeepSeek” unveiled this technology last October, describing it as capable of revolutionizing the processing of large and complex documents by using visual perception as a method of compressing texts. According to the company’s statements, this method is capable of reducing the number of text tokens by 7 to 20 times.
However, the recent study, entitled “Visual Virtue or Linguistic Crutch? An In-depth Look at DeepSeek-OCR,” concluded that the technology’s performance was erratic and that it relied heavily on what is known as “linguistic priors,” that is, the methods that the model derived from massive amounts of text, rather than the true visual understanding that the technology claimed to provide.
The researchers described the performance metrics announced by “DeepSeek” as “misleading,” noting that the accuracy of answering visual questions dropped to about 20% when additional text that could affect the conclusion was added, compared to more than 90% for traditional AI models.
The researchers explained that AI models still face fundamental difficulties in dealing with long documents or extended dialogues, a problem that companies and research centers around the world are seeking to solve. However, the results of the study raise fundamental questions about whether current visual compression technologies truly represent a practical way to overcome these difficulties.
“DeepSeek” did not immediately respond to a request for comment, but some computer science experts believed that the technology is not necessarily a failure, but rather a double-edged sword. Li Bujie, who holds a PhD in computer science from the University of Science and Technology of China and currently runs a startup in Beijing, said that relying on acquired knowledge may be useful when dealing with unclear manuscripts, but it may turn into a weakness when reading clearly printed texts, adding: “It can be said that this method has its advantages and disadvantages at the same time.”
This study demonstrates that the competition to improve AI capabilities in understanding long texts is ongoing, and that there are still no “magic solutions” that work for all cases, which is driving researchers to explore alternative and more effective strategies. (Al-Youm Al-Sabea)
