Confidence Calibration

Platt Scaling is a parametric method that fits a logistic regression model to the outputs of a classifier to map its scores to calibrated probabilities. It is most effective for binary classification tasks.

• Mechanism: Takes raw classifier scores (e.g., logits) and applies a sigmoid transformation with learned parameters (scale and bias).
• Use Case: Commonly used to calibrate Support Vector Machines (SVMs) and neural networks. It requires a separate, held-out validation set for training the scaling parameters to avoid overfitting.
• Limitation: Assumes the distribution of scores follows a sigmoidal shape, which may not hold for all models or datasets.

Isotonic Regression is a non-parametric calibration method that learns a piecewise constant, non-decreasing function to map uncalibrated scores to calibrated probabilities. It is more flexible than Platt Scaling.

• Mechanism: Does not assume a specific functional form (like sigmoid). It finds a stepwise function that minimizes the squared error between predicted scores and true binary outcomes, subject to a monotonicity constraint.
• Use Case: Effective for problems where the relationship between scores and true probabilities is complex and non-sigmoidal. Requires more calibration data than parametric methods to avoid overfitting.
• Consideration: Can be prone to overfitting on small datasets due to its flexibility.

Temperature Scaling is a simple, single-parameter extension of Platt Scaling designed for modern neural networks with multiple output classes. It is the most common method for calibrating large language models.

• Mechanism: Introduces a temperature parameter T > 0 to soften the final softmax output: softmax(logits / T). A T > 1 flattens the distribution (increases uncertainty), while T < 1 sharpens it.
• Optimization: The optimal T is found by minimizing the Negative Log Likelihood (NLL) on a validation set. It does not change the model's predicted class ranking, only the confidence estimates.
• Advantage: Preserves the model's accuracy while improving calibration, and is computationally very efficient.

Expected Calibration Error (ECE) is the primary quantitative metric for evaluating the quality of a model's calibration. It measures the difference between a model's confidence and its empirical accuracy.

• Calculation:
  1. Bin predictions into M equally spaced intervals based on their confidence score (e.g., [0.0, 0.1], [0.1, 0.2]).
  2. For each bin, compute the average confidence and the actual accuracy (fraction of correct predictions).
  3. ECE is a weighted average of the absolute difference between confidence and accuracy across all bins.
• Interpretation: A perfectly calibrated model has an ECE of 0, meaning its 70% confident predictions are correct 70% of the time. High ECE indicates miscalibration.
• Variants: Maximum Calibration Error (MCE) looks at the worst-case bin, while Static Calibration Error (SCE) extends ECE to multi-class settings.

Bayesian methods for calibration treat model parameters as distributions rather than point estimates, inherently capturing predictive uncertainty. This leads to better-calibrated outputs, especially with limited data.

• Core Principle: Instead of outputting a single probability, the model outputs a distribution over probabilities. The mean of this distribution is the predicted probability, and its variance represents epistemic uncertainty.
• Techniques: Include Monte Carlo Dropout (using dropout at inference to sample multiple predictions) and Deep Ensembles (training multiple models with different initializations).
• Benefit: These methods can distinguish between aleatoric uncertainty (noise inherent in the data) and epistemic uncertainty (model's lack of knowledge), providing richer, more honest confidence estimates.

Vector Scaling and Matrix Scaling are generalizations of Platt Scaling designed for multi-class classification problems, offering more flexibility than Temperature Scaling.

• Vector Scaling: Learns a separate scale parameter for each class and a single bias parameter. Transformation: softmax(W * logits + b), where W is a diagonal matrix.
• Matrix Scaling: Learns a full weight matrix W and bias vector b, allowing for interactions between classes: softmax(W * logits + b). This is the most flexible parametric form.
• Trade-off: Increased flexibility (Matrix > Vector > Temperature) allows for better calibration on complex miscalibration patterns but requires more calibration data and is more prone to overfitting. Temperature scaling is often preferred for its simplicity and robustness.

The overarching process of identifying when a generative model produces factually incorrect, nonsensical, or unsupported content not grounded in its source data or general knowledge. It is the primary problem that confidence calibration aims to solve by providing reliable probability scores.

• Core Goal: Flag outputs that are fabrications.
• Methods Include: Factual consistency checks, NLI, contradiction detection.
• Calibration's Role: A well-calibrated confidence score acts as a direct, quantitative signal for a detection system.

An evaluation method that verifies whether the claims or statements in a generated text are logically supported by a provided source document. It's a key downstream application of a calibrated model.

• Process: Compare each atomic claim in the output against the source.
• Relationship to Calibration: A model with miscalibrated confidence may claim high probability for statements that a consistency check would easily disprove, creating a false sense of security.

A method using pre-trained NLI models to classify the relationship between a generated claim and a source text as entailment, contradiction, or neutral. This provides a probabilistic assessment of factual support.

• Mechanism: The NLI model outputs a probability distribution over the three classes.
• Calibration Link: The confidence scores from the NLI model itself must be calibrated to ensure its contradiction/entailment predictions are trustworthy for automated detection pipelines.

Uses a classifier model (e.g., a cross-encoder) to directly judge the truthfulness of a claim given a context, outputting a probability score. This is a direct form of a verifier model.

• Contrast with Generative Verification: Classifies rather than generates justifications.
• Primary Output: A confidence score (e.g., 0.95 for 'supported').
• Critical Dependency: The utility of this score is entirely dependent on its calibration. An uncalibrated verifier is unreliable for decision-making.

A separate, often smaller model trained to evaluate the factuality, correctness, or safety of outputs from a primary language model. It acts as an external auditor.

• Function: Takes a (claim, context) pair and outputs a scalar score or classification.
• Calibration Imperative: For a verifier to be actionable—e.g., to set a threshold for automatic rejection of outputs—its scores must be calibrated to reflect true correctness likelihoods.

A quantitative metric measuring the proportion of factual claims within a model's output that are incorrect. It is a key performance indicator for hallucination detection systems.

• Calculation: (Number of False Claims) / (Total Claims Assessed).
• Connection to Calibration: A model with perfect calibration would have a Factual Error Rate for claims scored above a threshold p that is exactly (1 - p). For example, for all claims scored with 0.9 confidence, only 10% should be errors.