Uncertainty Quantification for Large Language Models: Making AI Responses More Reliable

This study focuses on Uncertainty Quantification (UQ) methods for Large Language Models (LLMs), aiming to address the hallucination problem of LLMs, evaluate output confidence, and provide methodological support for improving the reliability of AI systems. The study systematically analyzes and compares various UQ methods, covering background, methods, experiments, findings, and practical recommendations, offering important references for building more reliable AI systems.

Reliability Challenges of LLMs

Mainstream LLMs are based on the Transformer architecture, with a training objective of generating the most probable next word. They cannot distinguish between 'knowing' and 'not knowing', and tend to fabricate seemingly reasonable answers (hallucinations). Traditional Softmax probabilities are not proportional to correctness.

Value of UQ

Effective UQ methods can enable risk warning (prompting manual review), active learning (guiding data collection), decision support (adjusting downstream strategies), and user trust (establishing reasonable expectations), which are particularly crucial in high-risk scenarios such as healthcare and law.

UQ Classification Framework

LLM uncertainty is divided into two categories:

• Epistemic uncertainty: Arises from insufficient knowledge and can be improved through data/model scale;
• Aleatoric uncertainty: Arises from task ambiguity and cannot be eliminated (e.g., sentiment judgment of neutral text).

Evaluated UQ Methods

1. Probabilistic methods: Softmax + temperature scaling, Top-p sampling analysis;
2. Consistency methods: Multiple sampling consistency (Self-Consistency), semantic similarity clustering;
3. Verification methods: Self-verification, Chain-of-Thought confidence;
4. Model methods: Ensemble methods, Bayesian neural network approximations (e.g., Monte Carlo Dropout).

Datasets

Covers four types of Q&A tasks: factual questions, reasoning questions, open-ended questions, adversarial questions (inducing hallucinations).

Evaluation Metrics

• Calibration: Expected Calibration Error (ECE) measures the consistency between confidence and accuracy;
• Ranking ability: AUROC evaluates the ability to distinguish between correct and incorrect answers;
• Rejection performance: The magnitude of accuracy improvement at different confidence thresholds.

Key Findings

1. No single optimal method: Probabilistic methods are efficient but difficult to calibrate; consistency methods are reliable but costly; verification methods are flexible but require careful prompting;
2. Self-Consistency performs prominently: Balances effectiveness and cost in most tasks, and is more accurate when combined with semantic similarity;
3. Chain-of-Thought improves UQ quality: Confidence changes in reasoning steps can identify the model's 'lost' points;
4. Significant domain specificity: Uncertainty in factual questions comes from knowledge gaps; in reasoning questions from logical complexity; in open-ended questions from ambiguous answer boundaries.

1. Resource-constrained scenarios: Use calibrated Softmax probabilities as a lightweight solution;
2. Critical application scenarios: Adopt the Self-Consistency method (5-10 samples to calculate consistency scores);
3. Hybrid strategy: Probabilistic methods for quick screening of uncertain answers, consistency methods for fine evaluation of boundary cases;
4. Dynamic thresholds: Use high thresholds for high-risk scenarios (e.g., healthcare), and relax thresholds for creative scenarios.

Current Limitations

• Limited evaluation scope (only Q&A tasks, not covering code generation, long texts, etc.);
• High computational cost for some methods (e.g., ensemble, multiple sampling);
• Lack of causal analysis (only correlation).

Future Directions

• Fine-grained UQ (locating confidence of specific segments in answers);
• Multimodal UQ (cross-modal confidence for vision-language models);
• Adaptive UQ (automatically adjusting strategies);
• UQ in human-AI collaboration (effectively communicating uncertainty to users).