Uncertainty Estimation & Confidence Calibration¶

~2 минуты чтения

Предварительно: Фреймворки оценки LLM

Source: DNV Technology Insights "Practical Guide to Uncertainty Estimation" (2025), ArXiv "Uncertainty Quantification and Confidence Calibration" (2025), ICML 2025 Workshop

Концепция¶

Зачем UQ для LLMs: - High-stakes domains: healthcare, legal, finance - LLMs generate plausible-sounding but incorrect outputs - Need to know WHEN to trust model outputs

Types of Uncertainty:

Type	Description	Example
Epistemic	Knowledge gap (reducible with more data)	"What are 2025 emission standards?" for 2023-trained model
Aleatory	Inherent ambiguity (irreducible)	"Is this pump safe?" without context
Model-Intrinsic	Architecture/training uncertainty	Different valid answers from sampling
Content	Outside training distribution	Domain-specific questions

UQ Methods by Model Access¶

Level	Access	Methods
Black-box	API only	Sampling, verbalization, external validators
Grey-box	Probabilities/embeddings	Logit analysis, semantic clustering
White-box	Full model	BNN, MC Dropout, internal activations

Sampling-Based Uncertainty¶

class SamplingUncertainty:
    """Estimate uncertainty via multiple samples"""

    def __init__(self, model, n_samples=10, temperature=0.7):
        self.model = model
        self.n_samples = n_samples
        self.temperature = temperature

    def estimate_uncertainty(self, prompt):
        """Generate multiple responses and measure consistency"""
        responses = []

        for _ in range(self.n_samples):
            response = self.model.generate(
                prompt,
                temperature=self.temperature
            )
            responses.append(response)

        # Semantic clustering of responses
        embeddings = self.model.embed(responses)
        clusters = self.cluster_responses(embeddings)

        # Uncertainty metrics
        metrics = {
            "num_clusters": len(clusters),
            "largest_cluster_ratio": max(len(c) for c in clusters) / self.n_samples,
            "entropy": self.compute_cluster_entropy(clusters),
            "semantic_diversity": self.compute_diversity(embeddings)
        }

        # High uncertainty = many clusters, low largest cluster ratio
        is_uncertain = (
            metrics["num_clusters"] > 3 or
            metrics["largest_cluster_ratio"] < 0.5
        )

        return {
            "uncertainty_metrics": metrics,
            "is_uncertain": is_uncertain,
            "responses": responses
        }

Calibration Methods¶

Post-hoc Calibration:

class CalibratedConfidence:
    """Calibrate confidence scores post-hoc"""

    def __init__(self, calibration_data):
        self.calibrator = self.fit_calibrator(calibration_data)

    def fit_calibrator(self, data):
        """Learn mapping: model_confidence -> actual_accuracy"""
        from sklearn.isotonic import IsotonicRegression

        confidences = [d[2] for d in data]
        accuracies = [1.0 if d[3] else 0.0 for d in data]

        calibrator = IsotonicRegression(out_of_bounds='clip')
        calibrator.fit(confidences, accuracies)
        return calibrator

    def calibrate(self, confidence_score):
        return self.calibrator.predict([[confidence_score]])[0]

Интервью вопросы¶

Q: Что такое epistemic vs aleatory uncertainty?

A: Epistemic -- uncertainty из-за недостатка знаний (reducible с большим data). Aleatory -- inherent uncertainty из-за ambiguity/randomness (irreducible). Пример epistemic: "2025 emission standards" для модели обученной до 2023. Пример aleatory: "Is this pump safe?" -- зависит от условий не указанных в question.

Q: Как оценить uncertainty с black-box API доступом?

A: (1) Sampling-based: generate multiple responses, cluster by semantics, measure diversity. (2) Verbalized confidence: ask model to express its uncertainty. (3) External validators: use separate model/classifier. (4) Post-hoc calibration: learn mapping от model confidence к actual accuracy.

Q: Что такое calibrated confidence?

A: Calibrated confidence = confidence score matches actual accuracy. Если модель говорит "90% confident" -- она должна быть права в 90% случаев. LLMs часто overconfident. Calibration methods: Isotonic Regression, Platt Scaling, Temperature Scaling.