What is calibration in the context of probabilistic outputs from language models, and how can you calibrate them?

• A. Calibration adjusts the model's architecture.
• B. Calibration aligns predicted probabilities with empirical accuracy; calibrate with temperature scaling, isotonic regression, or Platt scaling on a held-out set.
• C. Calibration makes the model generate more creative text.
• D. Calibration only affects training speed.

Answer: (B)

Calibration is about making the model’s probability outputs reflect reality: when the model assigns a probability to a token or outcome, that probability should match how often that token actually appears in similar situations. If a token is labeled with 0.7 confidence, it should appear about 70% of the time across many predictions. Without calibration, those probabilities can be overconfident or underconfident, which can mislead decision points that rely on those numbers.

To calibrate, you use a held-out calibration set and fit a simple mapping from the model’s raw scores to true probabilities. Temperature scaling adjusts the logits by a single scalar temperature before applying softmax; this preserves the ranking of options but changes the confidence level, typically making outputs more or less confident to better align with observed frequencies. Isotonic regression is a flexible, nonparametric approach that learns a monotone mapping from uncalibrated scores to probabilities, which can capture complex relationships but needs enough data to avoid overfitting. Platt scaling fits a logistic regression model on the uncalibrated scores (or logits) to produce calibrated probabilities, effectively transforming the outputs with a sigmoid-like function.

All of these methods rely on evaluating the model on a separate calibration set and checking how well the predicted probabilities match actual outcomes, often using reliability diagrams, Brier score, or expected calibration error. The key idea is choosing a calibration method that provides a reliable, monotonic mapping from the model’s scores to well-calibrated probabilities, improving the alignment between stated confidence and empirical truth.