Uncertainty Quantification for Language Models: A Suite of Black-Box, White-Box, LLM Judge, and Ensemble Scorers

Technical Summary: Uncertainty Quantification for Language Models

Problem Statement
Large Language Models (LLMs) are increasingly deployed in high-stakes domains such as healthcare and finance, where factual accuracy is critical. A primary risk in these applications is "hallucination," defined here as the generation of content that sounds plausible but is factually incorrect. While pre-deployment evaluations comparing generated text to ground truth are common, they do not support real-time monitoring of deployed systems. Furthermore, many existing detection methods rely on access to source content or ground truth databases, which are unavailable in "closed-book" settings. The paper addresses the need for a versatile, closed-book framework that can compute response-level confidence scores at generation time to detect hallucinations without access to reference answers or external databases.

Methodology
The authors propose a unified framework that adapts existing Uncertainty Quantification (UQ) techniques into standardized confidence scores ranging from 0 to 1 (where 1 indicates high confidence). The framework categorizes scorers into three distinct types and introduces a tunable ensemble approach:

1. Black-Box UQ Scorers: These methods exploit the stochastic nature of LLMs by generating multiple candidate responses ($m$) from the same prompt and comparing them to the original response. The paper implements and normalizes several approaches:

   • Exact Match Rate (EMR): Measures the proportion of candidates identical to the original response.
   • Non-Contradiction Probability (NCP): Uses a Natural Language Inference (NLI) model to calculate the probability that candidates do not contradict the original response.
   • BERTScore (BSC): Computes the average F1 score based on contextualized word embeddings between the original and candidate responses.
   • Normalized Cosine Similarity (NCS): Measures semantic similarity using sentence transformers, normalized to the [0, 1] range.
   • Normalized Semantic Negentropy (NSN): A normalized version of Semantic Entropy that clusters responses by mutual entailment to measure response volatility.

2. White-Box UQ Scorers: These leverage internal token probabilities from the LLM's generation process:

   • Length-Normalized Token Probability (LNTP): The geometric mean of token probabilities, normalized for response length.
   • Minimum Token Probability (MTP): The lowest probability among all tokens in the generated response.

3. LLM-as-a-Judge Scorers: A separate LLM is prompted to evaluate the factual correctness of a question-response pair, outputting a score on a 0–100 scale which is then normalized to [0, 1].

4. Tunable Ensemble Approach: The core contribution is a weighted average of the individual scorers described above. The weights ($w$) are optimized using a user-provided set of graded LLM responses (where ground truth labels are known) and a classification objective function (e.g., AUROC or F1-score). This allows the ensemble to be customized for specific use cases, LLMs, and datasets.

Key Contributions

• Standardized Framework: The paper outlines a method to transform diverse UQ techniques into a consistent, response-level confidence metric suitable for real-time, closed-book hallucination detection.
• Open-Source Toolkit: The authors release uqlm, a Python package providing ready-to-use implementations of all proposed scorers and the ensemble tuning mechanism.
• Extensibility: The ensemble design is modular, allowing practitioners to incorporate new scoring methods as research evolves.
• Empirical Insights: The study provides data-driven guidance on scorer selection, such as the diminishing returns of sampling additional candidate responses and the correlation between a model's generation accuracy and its effectiveness as a judge.

Experimental Results
The authors evaluated the framework across six benchmarks (GSM8K, SVAMP, CSQA, AI2-ARC, PopQA, NQ-Open) using four LLMs (GPT-4o, GPT-4o-mini, Gemini-2.5-Flash, Gemini-2.5-Flash-Lite).

• Ensemble Performance: The tunable ensemble consistently outperformed individual components in 20 out of 24 LLM-dataset scenarios for AUROC and 17 out of 24 for F1-score.
• Scorer Variability: The performance of individual scorers varied significantly by dataset. NLI-based black-box scorers (NSN and NCP) generally outperformed other black-box methods.
• Judge Selection: A model's accuracy on a specific dataset positively correlated with its performance as a judge for that same dataset. For instance, Gemini-2.5-Flash excelled as a judge on math benchmarks, while GPT-4o performed best on short-answer benchmarks.
• White-Box vs. Black-Box: White-box scorers (LNTP, MTP) performed approximately equally to each other but generally underperformed compared to the best black-box or ensemble methods in many scenarios, particularly when token probabilities were not the primary signal of uncertainty.
• Filtering Efficacy: Applying confidence thresholds to filter low-confidence responses resulted in a monotonic increase in model accuracy. For example, filtering GPT-4o responses on GSM8K with the ensemble scorer raised accuracy from a baseline of 0.55 to 0.93 at a threshold of 0.6.

Significance and Claims
The paper claims that customized hallucination detection strategies, specifically through tunable ensembles, offer a practical and effective solution for improving the reliability of LLMs in production. The authors emphasize that there is no "one-size-fits-all" solution; rather, the optimal approach depends on the specific LLM, dataset, and operational constraints (e.g., latency, API access to token probabilities).

The authors maintain a modest stance regarding limitations, noting that their findings may not generalize to long-form generation or code generation without modification. They also caution that confidence scores reflect model uncertainty rather than ground-truth correctness, advising against their use as absolute decision guarantees in high-risk domains without human review. The work serves as a practical toolkit for practitioners to implement model-agnostic, real-time hallucination detection.