Supervised LLM Uncertainty Quantification

• The paper presents a framework that integrates Bayesian ensembles with LoRA adapters to mitigate overconfidence and calibrate LLM predictions.
• Supervised probes leveraging hidden activations and auxiliary discriminators yield robust uncertainty scores and improved calibration in both classification and generative tasks.
• Advanced techniques like prompt Bayesianization and evidential distillation offer efficient trade-offs between performance and inference cost in uncertainty quantification.

Supervised LLM uncertainty quantification encompasses a family of methodologies for assessing, diagnosing, and calibrating the confidence of LLM predictions in the presence of labeled data, typically through leveraging Bayesian, ensemble, or supervised discriminative techniques. These approaches not only seek to characterize predictive uncertainty—both aleatoric (data-intrinsic) and epistemic (model-based)—but also enable fine-grained detection of overconfidence, domain shift, and hallucination in the context of LLMs fine-tuned for downstream tasks or specialized domains.

1. Bayesian, Ensemble, and Posterior Approximation Frameworks

A dominant paradigm for supervised LLM uncertainty quantification is the Bayesian approximation of the predictor posterior via parameter-efficient deep ensembles. In low-rank adaptation (LoRA) settings, a fixed pre-trained LLM backbone with parameters $\omega_\mathrm{pre}$ is augmented with trainable low-rank adapters $(A, B)$, producing the parameterization

$W = W_\mathrm{pre} + BA$

with $B\in\mathbb{R}^{d\times r}$, $A\in\mathbb{R}^{r\times k}$, $r\ll\min(d,k)$ (Balabanov et al., 2024). Supervised fine-tuning is formulated as minimizing the variational KL objective

$$\min_{q(\theta)} \mathrm{KL}(q(\theta) \| p(\theta)) - \mathbb{E}_{q(\theta)} [\log p(t|s, \theta)]$$

where $$p(\theta)=\mathcal{N}(\theta; \omega_\mathrm{pre}, \lambda^{-1}I)$$ is a Gaussian prior.

A practical Bayesian posterior approximation utilizes M independent LoRA-fine-tuned models (ensemble members) $\{\theta_m\}_{m=1}^M$, each initialized from the same backbone but differentiated by seed and data order. This ensemble constitutes a variational family $\{p(t \mid s, \theta_m)\}$ for uncertainty computation.

Uncertainty decomposition follows from the predictive posterior: $$p(t^* \mid s^*, D) \approx \frac{1}{M} \sum_{m=1}^M p(t^* \mid s^*, \theta_m)$$ with Shannon entropy as total uncertainty and ensemble-based mutual information for epistemic component: $$\mathrm{MI}(\theta, t^* \mid s^*, D) = H_\mathrm{ens} - \frac{1}{M} \sum_{m=1}^M H_\mathrm{single}(\theta_m)$$ where $$H_\mathrm{ens} = -\sum_c p(t^*=c|s^*,D)\log p(t^*=c|s^*,D)$$, and $H_\mathrm{single}(\theta_m)$ is the entropy of the prediction from each ensemble member.

Empirical results on supervised multiple-choice tasks (CommonsenseQA, MMLU-SS, MMLU-STEM) demonstrate that LoRA ensembles suppress overfitting-induced overconfidence and provide robust signals of domain shift and data ambiguity (Balabanov et al., 2024).

2. Supervised Probes and Auxiliary Discriminators

Supervised discriminative approaches train lightweight models on top of hidden activations or attention-derived features to deliver uncertainty scores tightly correlated with correctness. Examples include:

• Bayesian Linear Lens: Layerwise Bayesian linear probes are trained to predict the residual change in hidden representations between LLM layers. Posterior predictive variances and log-likelihood ratios are then sparsely aggregated via elastic-net logistic regression to derive global uncertainty scores. This approach outperforms maximum-softmax-probability on MMLU for AUROC and expected calibration error (Dakhmouche et al., 5 Oct 2025).
• Feature Gap Measurement: Token-level epistemic uncertainty is formally linked to the KL divergence between the true and predicted token distributions, upper-bounded by the feature gap in hidden representations. Proxy directions (context reliance, comprehension, honesty) are elicited via contrastive prompting, and a small number of labeled samples suffice to calibrate a linear combination yielding a robust and data-efficient supervised UQ metric (Bakman et al., 3 Oct 2025).
• Hidden-Activation Regression: Simple regressors (e.g., random forests, MLPs) take as input concatenated grey-box features (token probabilities, entropy summaries) and white-box features (selected hidden activations), with the regression target being the scoring function between model output and ground truth. This framework achieves high AUROC and calibration across base and OOD domains (Liu et al., 2024).

3. Supervised Hallucination and Outlier Detection

For generative factual tasks requiring fine-grained detection of unsupported or hallucinated claims, specialized supervised modules have been proposed:

• Transformer UQ Heads: Lightweight transformer modules are trained on per-claim or per-token attention patterns and probability features, with labels constructed via external LLM adjudication (e.g., GPT-4o). These modules achieve state-of-the-art claim-level PR-AUC in both in-domain and out-of-domain hallucination detection, without requiring modification of the base LLM (Shelmanov et al., 13 May 2025).
• Robust Uncertainty for Trap Questions: For adversarial prompts designed to elicit hallucinations, uncertainty quantification is enhanced by decomposing emissions into facts and tailoring scoring rules to each fact type (e.g., refusal, correction, or falsehood). This method relies on external LLMs for fact decomposition and supervised classifiers for output type assignment and produces substantial ROC-AUC improvements over whole-text entropy or length-normalized entropy (Zhang et al., 1 Jan 2026).

Method/Class	Uncertainty Mechanism	Output Granularity
LoRA Ensemble Bayesian	Entropy+MI across models	Instance/prediction
Linear Probes/BLL	Posterior regression & ratios	Global per-instance
Feature Gaps	Hidden-representation distance	Per-instance/task-level
UQ Transformer Heads	Attention+prob for claims/tokens	Per-claim/per-token
Fact-aware Robust UQ	Supervised fact decomposition	Per-fact/per-response

4. Supervised Uncertainty Calibration and Evaluation

Calibration techniques in the supervised regime include temperature scaling, reliability diagrams, and direct minimization of proper scoring rules (NLL, Brier score). Expected Calibration Error (ECE) is standard for quantifying the difference between empirical accuracy and model confidence across stratified bins. Supervised UQ methods utilizing held-out labeled data typically outperform all unsupervised baselines on AUROC, ECE, and robustness under domain shift (Zhang et al., 5 Dec 2025, Liu et al., 2024, Dakhmouche et al., 5 Oct 2025).

Conformal prediction, using nonconformity scores based on softmax probabilities, offers finite-sample guarantees for coverage rate at a user-specified error $\alpha$, with average set size acting as an uncertainty surrogate. In supervised benchmarking, accuracy and uncertainty do not always co-vary, with larger or instruction-fine-tuned models showing greater conservatism (larger prediction sets) even as accuracy increases (Ye et al., 2024).

5. Advanced UQ: Prompt/Post-hoc Bayesianization and Knowledge Distillation

Recent methods extend supervised UQ for LLMs through:

• Textual Parameter Bayesianization: Treating pipeline prompts themselves as latent variables, posterior inference via Metropolis-Hastings with LLM proposals enables Bayesian uncertainty over textual parameters even in proprietary black-box pipelines. Empirically, this approach provides superior calibration (lower semantic ECE) and abstention quality relative to traditional data augmentation (Ross et al., 11 Jun 2025).
• Efficient Evidential Distillation: By distilling uncertainty-aware teacher ensembles (e.g., Bayesian Prompt Ensembles) into LoRA-tuned student models with Dirichlet or softmax heads, it is possible to match or outperform the teacher’s UQ performance in a single student forward pass, reducing inference cost by an order of magnitude (Nemani et al., 24 Jul 2025).

6. Limitations, Open Problems, and Empirical Observations

Although the supervised LLM uncertainty quantification landscape offers robust, calibrated methods, several challenges and observations merit consideration:

• Data availability: Supervised calibrators require labeled data or reliable surrogate measures, which may not exist for every target domain (Liu et al., 2024).
• Explanatory power: Methods using hidden-activation or attention-based features can be highly predictive but may offer less interpretability than explicit Bayesian posteriors or conformal methods.
• Cross-domain generalization: Feature gap methods and transformer UQ heads demonstrate strong robustness in OOD tests, but all approaches may be vulnerable to distribution shift if the feature or prompt space is inadequately covered (Bakman et al., 3 Oct 2025, Shelmanov et al., 13 May 2025).
• Overconfidence detection: Methods that explicitly target epistemic uncertainty (e.g., mutual information, feature gaps, fact-aware refinement) are more effective under adversarial or open-domain scenarios, where naively high softmax probabilities can mask outright error (Zhang et al., 1 Jan 2026).
• Practical tradeoffs: There is a persistent latency vs. UQ quality tradeoff. LoRA ensembles and prompt Bayesianization are relatively expensive at inference time, while knowledge-distilled or per-claim head approaches are lightweight (Balabanov et al., 2024, Nemani et al., 24 Jul 2025, Shelmanov et al., 13 May 2025).

7. Practical Guidelines

• Use ensemble or Bayesian-inspired approaches with LoRA or prompt diversity for explainable, instance-level UQ.
• Integrate discriminative or regression-based calibrators on hidden/attention features for maximal AUROC and calibration.
• Distinguish uncertainty estimation (discrimination of correct/wrong) from calibration (probabilistic validity), and optimize each if labeled data permits.
• For hallucination detection, employ supervised claim-level heads or fact-aware scoring to explicitly address non-canonical prompts.
• In operational deployments, supplement softmax-derived confidence with entropy, mutual information, and post-hoc calibration; report ECE, AUROC, and reliability diagrams (Zhang et al., 5 Dec 2025, Ye et al., 2024, Liu et al., 2024).

Supervised LLM uncertainty quantification is now a mature, multifaceted field, with methodologically grounded solutions for both fine-tuned classification and free-form generative tasks. Continued advances leverage the latent geometry of hidden representations, Bayesian posterior modeling (including over prompts and LoRA parameters), and supervised auxiliary heads for claim or fact-level confidence, all converging to provide actionable, well-calibrated signals critical for high-stakes deployment and research evaluation (Balabanov et al., 2024, Dakhmouche et al., 5 Oct 2025, Bakman et al., 3 Oct 2025, Nemani et al., 24 Jul 2025, Shelmanov et al., 13 May 2025, Zhang et al., 1 Jan 2026, Liu et al., 2024, Ye et al., 2024).