Calibration of language models: three methods to align confidence and accuracy

Language models often "lie" with their confidence: they say "I'm 90% certain," but are wrong half the time. Calibration is a set of techniques that eliminate this gap without changing the model's weights. The article covers three of the most common post-hoc approaches.

Why miscalibration is dangerous

A well-calibrated model means: if it assigns "70% probability" to an event, in reality such an event should occur in roughly 70% of similar cases. This sounds reasonable, but in practice most modern LLMs are overconfident — they systematically overstate their own accuracy, especially in scenarios far from the training data. The problem is not abstract. In products where decisions are made based on model probabilities — medical diagnostics, credit scoring, legal expertise, trading algorithms — incorrect 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 two ways. ECE (Expected Calibration Error) — a single numerical metric that averages the absolute deviation between predicted probabilities and actual frequencies across ranges. Reliability diagram — a visual tool: X-axis is predicted probabilities, Y-axis is actual accuracy. Perfect calibration is a straight diagonal line. Downward deviation means overconfidence; upward means underconfidence.

Three post-hoc calibration methods

All three approaches are applied after neural network training. They do not change weights but correct output probabilities at the inference stage using small additional training on validation data. This makes them convenient: no need for access to model training, no need for additional annotation.

• Platt Scaling — trains logistic regression on top of raw model logits. Fast and simple, requires little data. Assumes sigmoidal distortion — works well if the model uniformly over- or under-estimates confidence across the entire range. Weakness: handles nonlinear patterns poorly.
• Isotonic Regression — a non-parametric method with no assumptions about distortion shape. Builds a monotonic transformation function of arbitrary form, better handles complex nonlinear overconfidence patterns. Requires significantly more validation examples — otherwise risks overfitting.
• Temperature Scaling — the most popular method for LLMs. Divides all logits by a single learned constant T (temperature). T > 1 — probabilities "blur," the model becomes less categorical. T < 1 — confidence increases. One parameter, optimized for NLL on validation. Implementation takes about ten lines of code.

How to choose the right method

Temperature Scaling — the default choice for most LLM tasks. Works only with output probabilities — no access to weights needed. Selecting T requires only a few hundred examples. Does not reshape the probability distribution, only scales it.

Isotonic Regression should be considered when distortion is nonlinear and the validation sample is large — from a few thousand examples. It can catch complex patterns that Temperature Scaling would miss. Overfitting risk requires careful train/val/test splitting.

Platt Scaling occupies middle ground: slightly more flexible than Temperature Scaling and simpler than Isotonic Regression. Makes sense when there's insufficient data for Isotonic but more flexibility is needed than temperature provides.

A key limitation of all three approaches: a separate validation set is required that was not used in model evaluation. Calibrating on the same set is an error leading to overfitting of the calibration function.

What this means

For teams deploying LLMs to production, calibration is a mandatory step before deployment in any high-stakes scenarios. Temperature Scaling requires minimal effort and can significantly reduce the number of incorrectly confident predictions — without overfitting, without access to weights, without noticeable latency increase. This is one of those rare cases where a small technical improvement directly impacts user trust in the product.