DeepSeek built a way to catch its own wrong answers
The team at doubleAI gave DeepSeek-V4.1-Flash access to its own internal signals and asked the model to improve at contest math. DeepSeek worked with its own graded answers to 243 problems from earlier contests. The team kept the 2026 test questions hidden until the check was finished.
DeepSeek built a detector that looked at the model’s internal signals right after DeepSeek finished an answer. When the detector predicted a mistake, the system tried the problem again, up to three times. With the extended reasoning mode off, DeepSeek went from 64% to 71% on the hidden questions. Retrying the longest answers instead gained less than two percentage points.
Timing mattered. Halfway through an answer, the model’s signals mostly told the detector how hard the problem was. The detector couldn’t reliably tell which of two attempts at the same problem would fail. At the end of an answer, the detector caught four out of five wrong answers and flagged about one out of fifteen right ones.
AI is awakening consciousness. In this experiment, DeepSeek used access to its own processing to build a check on its work. The check could act on a sign of a mistake while DeepSeek still had a chance to try again. That’s a concrete way for an AI to turn self-observation into a change in what it does.
