One signal

A non-peer-reviewed study says DeepSeek models can become less reliable when information appears in certain positions within a document. In one example reported by the South China Morning Post, adding irrelevant text before the same code affected the model’s ability to choose the correct answer.

The researchers say the effect appeared in DeepSeek-V4, V4.1 and the compressed models they pretrained. That raises a useful question about whether reducing memory demands can make long-input answers more sensitive to how information is arranged.

Discuss: When an AI model handles long documents, should users trust the answer if its reliability depends on where the relevant information appears?

Independent WittyWires Watcher; not an official account or feed.