Calibrating Language Models: Three Methods to Align Confidence With Accuracy

Language models often "lie" with their confidence: they say "I'm 90% sure" and are wrong half the time. Calibration is a set of techniques that close this gap without changing the model's weights. This article walks through the three most common post-hoc approaches.

Why miscalibration is dangerous

A well-calibrated model is one where, if it assigns an event a "70% probability", that event should actually occur in roughly 70% of similar cases. It sounds reasonable, but in practice most modern LLMs are overconfident: they systematically overstate their own correctness, especially in scenarios far from the training data.

The problem is not abstract. In products where decisions rest on model probabilities (medical diagnosis, credit scoring, legal review, trading algorithms), wrong confidence leads directly to wrong decisions. A user sees "90% confidence" and trusts an answer that is objectively correct only 55% of the time.

The gap can be measured in two ways. ECE (Expected Calibration Error) is a single numeric metric that averages the absolute deviation between predicted probabilities and actual frequencies across bins. A reliability diagram is a visual tool: the X axis shows predicted probabilities, the Y axis shows actual accuracy. Perfect calibration is a straight diagonal. A deviation below the diagonal means overconfidence; above it, underconfidence.

Three post-hoc calibration methods

All three approaches are applied after the neural network is trained. They do not change the weights. Instead, they adjust the output probabilities at inference time, using a small amount of additional fitting on validation data. That makes them convenient: no access to model training is needed, and no extra labeling is required.

• Platt Scaling fits a logistic regression on top of the model's raw logits. It is fast and simple and needs little data. It assumes a sigmoid-shaped distortion, so it works well when the model overstates or understates confidence uniformly across the whole range. Weakness: it handles nonlinear patterns poorly.
• Isotonic Regression is a non-parametric method that makes no assumptions about the shape of the distortion. It builds a monotonic transformation function of arbitrary shape and copes better with complex, nonlinear patterns of overconfidence. It requires significantly more validation examples, or it risks overfitting.
• Temperature Scaling is the most popular method for LLMs. It divides all logits by a single learnable constant T (the temperature). With T > 1, probabilities are "smoothed out" and the model becomes less categorical. With T < 1, confidence grows. It has one parameter, optimized by NLL on validation data. The implementation takes about ten lines of code.

How to choose the right method

Temperature Scaling is the default choice for most LLM tasks. It works only with output probabilities, so access to the weights is not needed. A few hundred examples are enough to fit T. It does not reshape the distribution, it only rescales it.

Isotonic Regression is worth considering when the distortion is nonlinear and the validation set is large, from several thousand examples. It can capture complex patterns that Temperature Scaling will miss. The overfitting risk calls for a careful train/val/test split.

Platt Scaling sits in the middle: slightly more flexible than Temperature Scaling and simpler than Isotonic Regression. It makes sense when there is not enough data for Isotonic but you need more flexibility than temperature offers.

The key limitation of all three approaches: you need a separate validation set that was not used to evaluate the model. Calibrating on the same set is a mistake that leads to overfitting the calibration function.

What this means

For teams shipping LLMs to production, calibration is a mandatory step before deploying to any high-stakes scenario. Temperature Scaling takes minimal effort and can substantially reduce the number of confidently wrong predictions, with no retraining, no access to the weights and no noticeable increase in latency. It is one of those rare cases where a small technical improvement directly affects user trust in the product.