Concept认知谦逊测试 / 认知校准
epistemic calibration
A concept in AI evaluation referring to a model's ability to accurately perceive and express the boundaries of its own knowledge—acknowledging uncertainty when evidence is insufficient rather than producing overconfident but incorrect answers.
Timeline (last 90 days)
Oct 11
The epistemic calibration test examines whether a model can accurately perceive the boundaries of its own knowledge, whereas traditional LLM benchmarks such as MMLU and HumanEval typically use accuracy as the sole metric.
Unverified50%