Probability Calibration

Calibrated confidence means that when the system gives a score of 0.9, the probability of being correct should be close to 90%.

The probability figures provided by guided generative models are essentially 'well-constrained inferences' rather than directly verifiable truths.

A "well-calibrated" probability judgment means that when a model says something has a 70% chance of occurring, the actual occurrence rate of that event across many similar predictions is close to 70%.

Probability calibration refers to a post-processing technique that converts model output prediction scores into true probability estimates, with commonly used methods including Platt Scaling and Isotonic Regression.

Model calibration refers to the degree to which a model's output probabilities match its actual accuracy rate; for a well-calibrated model, if the probability it gives for a certain type of question is 0.8, then about 80% of such questions are indeed answered correctly.

Probability calibration is an evaluation dimension of classification models that is independent of accuracy.

Probability Calibration is a concept in machine learning for assessing the trustworthiness of model outputs.

Routing scenarios typically rely on probabilistic scores for hard-threshold decisions, but neural network softmax outputs often suffer from poor calibration, so in practice post-calibration methods such as temperature scaling and histogram binning are needed, along with ECE evaluation.

For PD models, calibration quality directly affects expected loss calculation; a model with a high AUC but poor probability calibration may lead to distorted provisioning.

Accuracy alone cannot support gating thresholds; the decision model must be well calibrated so that the output probability values reflect true occurrence rates.

The probability calibration of model outputs is often biased, especially on patterns that appear familiar within the training data distribution, where overly high confidence is given.

The original material did not mention whether Nimble has built-in calibration, so before actual integration it is best to first plot a reliability diagram with known samples to confirm whether the probability distribution is usable.

未校准的模型若直接套用固定阈值做自动决策，会系统性地高估或低估风险，因此上线前应绘制可靠性图并评估ECE等指标。

LLM输出的置信度来自logits或token概率，是模型内部的相对偏好，并不等同于事件发生的真实频率，即模型可能过度自信。

Conventional training (such as cross-entropy loss) only penalizes incorrect predictions without constraining probability calibration quality, causing models to be overconfident or overly conservative.

The probability reported by a model (e.g., 90%) refers to the model's confidence in its own classification judgment and has nothing to do with business urgency.

Probabilities can help manage the overall risk across a batch of requests, but cannot provide guarantees for any single request.

The study identifies a decision-time leakage problem: the controller peeks at future information that should not be available at the time of decision-making, which systematically overestimates the benefits brought by prediction.

Decision quality is highly dependent on the accuracy of the posterior probability P(y|x); model overconfidence leads to systematic bias in expected loss calculations, so it is recommended to first perform probability calibration such as Platt scaling or isotonic regression.