What DeepSeek model accuracy research actually shows

On October 12, 2026, the South China Morning Post reported that a ByteDance-led paper published in late September found that DeepSeek models could miss information depending on where it appeared in a document. Separate Nonobench scores measure nonogram solving, a different task—not that retrieval effect.

What the reported study says about retrieval

Position-sensitive retrieval means a model’s ability to find information can vary with that information’s location in a document. In the example described by the South China Morning Post, adding irrelevant text before the same code impaired DeepSeek’s ability to select the correct answer. The report linked memory-saving approaches for processing long inputs with a possible reduction in retrieval reliability; it gave no numerical effect size. (South China Morning Post’s report)

That is a specific reported weakness, not a general accuracy score. The paper’s reported scope was DeepSeek-V4, V4.1, and all compressed models its authors pretrained. The South China Morning Post said the paper had not been peer-reviewed as of its October 12, 2026, report.

What the Nonobench scores measure

A separate Nonobench v1.2 discussion reports results for two models on nonogram puzzles, which test puzzle-solving rather than retrieval from a document:

The discussion author describes hard mode as 10 random 20×20 puzzles, with one attempt per puzzle, and reports that every open-weight model scored 0/10. That hard-mode result is separate from the two scores in the table. The author says the runs used single prompts, without a harness or tool calls. (Nonobench v1.2 discussion)

The retrieval finding and the puzzle scores concern different tasks. The Nonobench results therefore neither measure nor confirm the reported position-sensitive retrieval effect.