Uncertainty quantification (UQ) for large language models is one of the more practically urgent problems in the field. When a model hallucinates confidently, downstream systems, whether clinical, legal, or scientific, have no principled way to detect the failure. Most existing UQ methods operate zero-shot: they generate a score correlated with correctness without requiring any labelled data. This paper, available at arXiv:2608.19323, makes a simple but underexplored observation: in practice, you always evaluate your model on a representative dataset before deployment anyway, so why not use that data to improve your confidence estimates?
The answer the authors give is a supervised classification framework that treats existing UQ scores as features rather than final outputs. The result is consistent improvement across models, datasets, and uncertainty scoring methods, at negligible additional computational cost.
Key Contributions
- A reframing of LLM confidence estimation as a supervised binary classification problem, predicting whether a given response is correct.
- A feature construction pipeline that combines existing zero-shot UQ scores (verbalized confidence, semantic entropy variants) with neighbourhood-based statistics derived from a reference set of labelled queries.
- Empirical demonstration that simple classifiers, specifically logistic regression and random forests, consistently outperform all zero-shot baselines on AUROC across four datasets and four models (GPT4.1, GPT4.1-mini, GPT4.1-nano, LLaMA4Maverick).
- Well-calibrated probability outputs, assessed via expected calibration error (ECE), with the method matching or beating temperature-scaled verbalized confidence in most settings.
Methodology
The setup is straightforward. Given a labelled evaluation dataset, the authors partition it into a training set and a reference set. For each query in the training set, they extract features in two ways. First, they run existing UQ methods, including verbalized confidence (P_true), semantic entropy, Laplacian entropy, and kernel entropy, to obtain scalar uncertainty scores. Second, they retrieve the k=20 nearest neighbours of the query from the reference set using embedding-space cosine similarity, then compute neighbourhood statistics: the average correctness of those neighbours and their similarity scores to the query.
These combined features are fed into either an l1-regularised logistic regression or a random forest classifier, with the best-performing model selected per task. The authors describe this as an Auto-ML procedure, though in practice it amounts to selecting between two model classes. The reference set itself is constructed using determinantal point processes to encourage diversity, which is a sensible choice for ensuring broad coverage of the query distribution.
The conceptual framing is also worth noting. The authors draw a distinction between heuristic scoring, which produces abstract scores intended to correlate with correctness, and confidence estimation, which targets calibrated probability estimates directly. They argue for the latter as the more useful quantity, since calibrated probabilities can be integrated into decision-theoretic frameworks in ways that raw heuristic scores cannot. This is a reasonable position, though the empirical gap between a well-calibrated heuristic and a true probability estimate may be small in many practical settings.
Results
The results in Table 1 are fairly compelling. Across all twelve dataset/model combinations tested with the GPT4.1 family, the Auto-ML method achieves the highest AUROC in eleven of twelve cases, with the single exception being marginal. The gains over the best zero-shot baseline range from modest (around 0.01 AUROC on SciQ with GPT4.1) to more substantial (around 0.05 on NQ with GPT4.1-nano). On ECE, the method consistently beats raw P_true and is competitive with temperature-scaled P_true, which is itself a supervised method requiring a scalar fit on held-out data.
A few patterns are worth highlighting:
- Semantic entropy-based methods perform poorly on multiple-choice datasets (CS_QA, SciQ), being outperformed by verbalized confidence alone. This aligns with known limitations of sampling-based methods when the answer space is constrained.
- APRICOT, a neural network-based confidence predictor, performs surprisingly poorly across the board. The authors attribute this partly to the absence of task-specific tuning, which is a fair caveat, but it also suggests that more complex models do not automatically win here.
- Results on LLaMA4Maverick (Table 2) show similar trends, confirming the method is not GPT-family specific.
The calibration results deserve separate attention. Temperature scaling of P_true (ts-P_true) is a strong and cheap baseline for ECE, and the Auto-ML method does not always beat it. However, the Auto-ML method achieves substantially better AUROC than ts-P_true in most cases, meaning it is simultaneously more discriminative and comparably calibrated. That combination is what you actually want in deployment.
Limitations and Implications
The authors are candid about two main limitations. The method requires task-specific classifiers, so there is no single model that transfers across domains out of the box. This is a real constraint, though as they note, domain-specific evaluation datasets are already standard practice before deployment. The second limitation is dependence on a sufficiently large reference set for meaningful nearest-neighbour retrieval. With very small datasets, the neighbourhood statistics will be noisy and may add little signal.
There are a few additional considerations the paper does not fully address. The choice of embedding model for nearest-neighbour retrieval is not ablated, and it is plausible that embedding quality has a significant effect on the neighbourhood features. The method also assumes that the evaluation dataset is representative of deployment queries; distribution shift between evaluation and production would degrade the neighbourhood features specifically, since retrieved neighbours would be less semantically relevant.
The broader implication is that the zero-shot framing that dominates UQ research may be unnecessarily restrictive. If labelled data is available, even a small amount, treating UQ as a learning problem rather than a scoring problem is likely to help. The authors' framework is modular enough that it could accommodate more expressive classifiers, and future work integrating retrieval-augmented features with lightweight neural predictors seems like a natural extension.
The paper also gestures toward multi-class correctness prediction as future work, moving beyond binary correct/incorrect labels to finer-grained quality assessments. This is an important direction; binary correctness is a coarse target that collapses a lot of meaningful variation in response quality, and richer labels would enable more nuanced abstention and routing strategies in production systems.
Overall, this is a practically motivated paper with clear empirical support for its central claim. The methodology is simple enough to implement quickly, the computational overhead is genuinely low, and the consistent gains across models and datasets make a credible case that supervised confidence estimation should be part of the standard UQ toolkit for LLM deployment.
Full paper: arXiv:2608.19323 | Code: github.com/layer6ai-labs/improved_llm_confidence