使用大语言模型生成置信度评分存在根本缺陷

大语言模型生成的置信度评分不应被用于实际应用，因为这些评分缺乏可信置信度的两个关键属性。1 可信的置信度评分必须反映预测所基于的信息量，同时需要根据实际结果进行校准，而LLM生成的置信度既不满足这两点。1

LLM生成的置信度存在多个可观察的问题。1 重复提出相同问题会导致置信度数字发生变化，改变表述方式同样会改变数值，而前文内容也会对结果造成影响。1 这些变化表明LLM的置信度本质上是文本生成的自然延续，而非对实际内容的真实测量。1 从技术角度看，LLM仅在"下一个文本是什么"这一问题上被预训练并校准，而非在"置信度与正确率相匹配"这一维度上校准。1

要构建可靠的置信度评分，无需依赖LLM，而应采用替代方案。1 该方案包含五个步骤：选择可观察的信息作为基础，从历史基数开始建立参考点，记录每个回复及其对应的结果，根据这些记录更新评分规则，定期验证校准情况。1

A critical examination of confidence scoring generated by large language models reveals fundamental flaws in their reliability and design.1 The core issue is that LLM-generated confidence scores represent merely a natural continuation of text rather than an actual measurement of predictive certainty.1

Trustworthy confidence scores require two essential attributes: they must reflect the volume of information underlying a prediction, and they must be calibrated against actual outcomes.1 LLM confidence scores fail on both counts. Observable problems demonstrate this instability—repeating the same question produces different numerical scores, altering phrasing changes the results, and preceding context influences the output.1 This inconsistency stems from a fundamental training mismatch: large language models are pre-trained to predict the next token accurately, not to align confidence levels with correctness rates.1

Rather than relying on LLM-generated confidence, an alternative approach builds calibrated scores through deliberate methodology.1 This five-step process begins by selecting observable information, establishing a baseline from historical data, documenting each response and its corresponding result, updating decision rules from these records, and periodically validating calibration accuracy.1 Such systems bypass the limitations of language models to produce confidence scores that genuinely reflect prediction reliability.