Metacognitive Prompting
A prompt engineering method that encodes the monitoring and regulation structure of human metacognition into a large model's reasoning process. It typically comprises five stages: comprehension, preliminary answer, criticism, decision, and confidence reporting, emphasizing the use of the criticism stage to catch errors missed in the first round.
Timeline (last 90 days)
The five-stage metacognitive prompt pipeline (understand, initial answer, critique, decide, confidence) raised the number of correctly answered tasks from 5/17 to 15/17 on a deterministic model.
After removing the gating condition from the metacognitive prompt pipeline, accuracy fell to zero, reverting to the 5/17 baseline.
The gated metacognitive pipeline rescued 10 wrong answers and broke 0 correct answers; the ungated version rescued 5 and broke 5 correct answers.
The gating rule is: only when the critique stage outputs a named defect does the decision stage make a modification; if the critique does not point out a specific error, the initial answer is returned unchanged.
Ablation experiments show that removing the understand stage only loses fixes for misreading-type tasks, removing the critique stage only loses fixes for rule-omission-type tasks, and the sum of the two gains equals the total gain with no overlapping attribution.
The confidence reported by the model is severely disconnected from actual accuracy: the 95% confidence tier had an actual accuracy of 0%, while the 90% and 70% tiers both had an actual accuracy of 100%.
Over-extrapolation (confidently answering questions that have no answer) is the only failure mode that passes through all five stages, because no step in the pipeline checks whether the question is underdetermined.
All Facts (7)
The five-stage metacognitive prompt pipeline (understand, initial answer, critique, decide, confidence) raised the number of correctly answered tasks from 5/17 to 15/17 on a deterministic model.
50%UnverifiedAfter removing the gating condition from the metacognitive prompt pipeline, accuracy fell to zero, reverting to the 5/17 baseline.
50%UnverifiedThe gated metacognitive pipeline rescued 10 wrong answers and broke 0 correct answers; the ungated version rescued 5 and broke 5 correct answers.
50%UnverifiedThe gating rule is: only when the critique stage outputs a named defect does the decision stage make a modification; if the critique does not point out a specific error, the initial answer is returned unchanged.
50%UnverifiedAblation experiments show that removing the understand stage only loses fixes for misreading-type tasks, removing the critique stage only loses fixes for rule-omission-type tasks, and the sum of the two gains equals the total gain with no overlapping attribution.
50%UnverifiedThe confidence reported by the model is severely disconnected from actual accuracy: the 95% confidence tier had an actual accuracy of 0%, while the 90% and 70% tiers both had an actual accuracy of 100%.
50%UnverifiedOver-extrapolation (confidently answering questions that have no answer) is the only failure mode that passes through all five stages, because no step in the pipeline checks whether the question is underdetermined.
50%