Accepted Papers (2026)
Archival
Christopher Kissling, Elena Merdjanovska, Alan Akbik. Self-Aware Knowledge Probing: Evaluating Language Models’ Relational Knowledge through Confidence Calibration
Knowledge probing quantifies how much relational knowledge a language model (LM) has acquired during pre-training. Existing knowledge probes evaluate model capabilities through metrics like prediction accuracy and precision. Such evaluations fail to account for the model’s reliability, reflected in the calibration of its confidence scores. In this paper, we establish a framework to evaluate the calibration of relational knowledge in a closed answer-set setting. We cover three modalities of model confidence: (1) intrinsic confidence, (2) paraphrase consistency and (3) linguistic confidence. Our extensive analysis of 21 causal and 6 masked language models reveals that most models, but especially those pre-trained with the masking objective, are overconfident. The best-calibrated scores come from confidence estimates that account for inconsistencies due to statement rephrasing. Moreover, calibration improves when increasing the number of answer options.
Chenyang Yu, Zifan Zeng, Qunli Zhang, Feng Liu. STEP: Soft Trajectory Estimation of Prediction Correctness via Per-Step Probability Scoring
Estimating the correctness of large language model (LLM) predictions without access to ground truth is critical for reliable deployment in high-stakes settings. Existing approaches, such as entropy and perplexity, operate solely on final output statistics, discarding the rich information encoded in intermediate reasoning steps. To address this limitation, we propose STEP (Soft Trajectory Estimation of Prediction correctness), a training-free framework that scores each reasoning step via lightweight forward passes to obtain a soft probability distribution $q_i$ over candidate answers, from which we derive a set of soft trajectory features capturing per-step confidence and its progression toward the final answer. Extensive experiments on various benchmarks demonstrate consistent and substantial improvements, particularly in AUROC, over output-based baselines and hard-indicator trajectory methods. Notably, Temporally-Weighted Consistency, which assigns higher importance to later reasoning steps where answer distributions have stabilized, contributes the largest gains, highlighting the importance of temporal structure in reasoning trajectories. These results further validate the advantage of preserving continuous probability signals over binary discretization.
Mario Sanz-Guerrero, Katharina von der Wense. Calibration as a First-Class Criterion in LLM Evaluation
Calibration of language models – the alignment between expressed or implicit confidence and empirical correctness – is a well-studied subfield within NLP. Methods to measure it already exist. The problem is adoption: outside this subfield, NLP research regularly introduces new models, datasets, and benchmarks without checking whether the model’s confidence scores are meaningful. We argue that this adoption gap is a major obstacle to trustworthy LLM evaluation. Miscalibration causes problems in two distinct areas: at deployment, where overconfident mistakes cause real harm, and inside the research pipeline, where methods like LLM-as-a-judge, synthetic data generation, and active learning rely on calibrated confidence without verifying it. Standard calibration metrics only require two inputs per example: a confidence score and a correctness judgment. Most benchmarks in use today already provide both, meaning calibration can be reported immediately. For open-ended generation, however, defining these two inputs is still an open challenge. We argue that each NLP subfield should pair its main performance metric with a calibration score and call for treating calibration as an essential property of every model rather than a niche topic.
Joshua Chaz Rivera, Vincent Gripon, Bastien Pasdeloup. Idiomatic Blind Spots in Discrete Diffusion Models
Discrete diffusion language models have emerged as a compelling alternative to autoregressive approaches, offering a range of structural advantages. This is due to their bidirectional attention allowing a global semantic view across intermediate denoising steps that enables flexible, order-agnostic generation. Despite this promise, they have received considerably less attention in ongoing research, and existing work has focused narrowly on mathematics and formal reasoning tasks. Their language comprehension capabilities, particularly for figurative language, remain entirely unevaluated, leaving open fundamental questions about how these models handle complex idiomatic structures, especially for languages other than English. We find that diffusion models do possess many strong idiomatic processing abilities, demonstrated through competitive token recovery in cloze tests and strong semantic ability from multiple-choice accuracy. However, this advantage fails to surface during inference as confidence-based demasking appears poorly suited for idioms, as early committed tokens override internal representations and discourage certain types of multi-token structures from generation, limiting their linguistic abilities.
Amit Ranjan Trivedi, Sruthikesh Surineni, Kyong Ho Lee, Mohit Mittal, Steven Li, Rex Hill. Trajectory-Triggered Sparse Collaborative Decoding
Uncertainty estimates in language models are typically computed locally, from the predictive distribution at an individual decoding step. In multi-step reasoning, however, the outcome of interest is often trajectory-level: whether the final answer produced after many intermediate steps will be correct. We show that these two notions of uncertainty can diverge sharply, and study what an uncertainty signal should look like when the prediction target is failure of the complete reasoning trajectory. We first quantify the limitation of local uncertainty for predicting final-answer failure. On GSM8K with a Qwen-1.5B draft model, token-level entropy predicts final incorrectness at AUC 0.531, negative log-likelihood at 0.528, and logit margin at 0.474. Short-window changes in these quantities remain between 0.498 and 0.503. A model can therefore remain locally confident while following an incorrect derivation, or become locally uncertain at a step from which it subsequently recovers. Next-token uncertainty does not directly measure the risk accumulated over a reasoning trajectory. The shape of the confidence trajectory is substantially more informative. We represent each observed prefix using approximately 85 statistics describing the evolution of entropy, likelihood, logit margin, and probability concentration, including means, extrema, percentiles, slopes, variances, temporal changes, and position-dependent behavior. A lightweight logistic predictor trained only on draft trajectories labeled by final correctness reaches AUC 0.864, while the strongest individual trajectory feature reaches 0.803. Its learned weights associate failure with declining logit-margin separation, increasing temporal volatility, and weaker top-k probability concentration. A complementary wavelet analysis finds elevated mid-frequency oscillation in incorrect trajectories, suggesting that persistent instability and unresolved competition between reasoning paths carry more information than isolated uncertainty spikes.
For an uncertainty estimate used during generation, full-trajectory discrimination can be misleading because the estimate must become reliable before the trajectory is complete. We therefore evaluate prefix-conditioned uncertainty. On GSM8K, the Qwen-1.5B predictor rises from AUC 0.60 after 16 tokens to 0.82 after 128 tokens and 0.86 on the complete trajectory. Across GSM8K drafts, AUC at 128 tokens ranges from 0.69 to 0.82. On MATH-500, full-trajectory AUC reaches 0.73-0.78, yet AUC at 128 tokens remains only 0.57-0.59. Harder symbolic problems accumulate evidence more slowly. These results motivate prefix-conditioned AUC as an operational measure of uncertainty quality for systems that must act during decoding.
We test whether this distinction matters in a downstream decision problem. In sparse collaborative decoding, a small draft model generates most of a reasoning trajectory, while a stronger target model is invoked only for short textual corrections when predicted failure risk crosses a threshold. Perturbation experiments reveal a fundamental timing tradeoff: early interventions have greater influence on the eventual trajectory, precisely when uncertainty estimates are least reliable, while later estimates are more accurate but have less ability to redirect generation.
Across five GSM8K draft-target configurations, trajectory-triggered intervention improves over the best fixed insertion schedule by 5.3-20.0 accuracy points and over draft-only decoding by as much as 43.0 points. On StrategyQA it improves over the best open-loop schedule by 5.0-8.3 points, demonstrating transfer beyond arithmetic reasoning. On MATH-500 it gains 7.0-13.5 points in successful configurations but loses when early trajectory prediction is weak. The failure cases expose an important practical distinction between retrospective and actionable uncertainty: a predictor can discriminate failures well after a trajectory is complete while remaining too unreliable at the prefix where a decision must be made. Target behavior introduces a second source of uncertainty. Corrective interventions that degrade an otherwise correct draft exhibit higher entropy and negative log-likelihood and lower maximum probability and logit margin than successful interventions. Consequently, stronger standalone target accuracy does not guarantee better corrective behavior. Reliable collaborative reasoning requires both an uncertainty estimate that identifies failing trajectories early enough to act and a target continuation that is itself sufficiently reliable.
Overall, our results suggest that uncertainty for multi-step language-model reasoning should be treated as a property of an evolving trajectory rather than an isolated token. Aggregate confidence dynamics reveal failure signals that local uncertainty largely misses, while prefix-conditioned evaluation distinguishes uncertainty that predicts eventual failure from uncertainty that is available early enough.
Susanna Rücker, Alexander Reinicke, Alan Akbik. SelectED: Selective LLM Refinement for Entity Disambiguation
Efficient architectures for Entity Disambiguation (ED) typically pair a dual encoder for candidate retrieval with a cross encoder for re-ranking. While large language models (LLMs) have recently emerged as powerful alternatives for the ranking stage, their application remains constrained by two primary factors: only a limited number of candidate entities can be provided within the model’s input context, and invoking LLMs for every entity mention leads to prohibitive computational cost.
In this work, we propose SelectED, a confidence-guided hybrid framework for efficient entity disambiguation. Our method uses a dual encoder as a fast primary model and selectively invokes an LLM only for uncertain mentions. A central focus is the systematic study of selective LLM refinement, including uncertainty estimation, thresholding, and prompt design choices for balancing accuracy and efficiency. Experiments on the ZELDA benchmark show that selective prompting improves accuracy by 4.3%, establishing a new best result among directly comparable ZELDA-trained systems while requiring LLM inference for only about 20% of mentions. Moreover, SelectED surpasses full LLM prompting, suggesting that selective refinement can avoid LLM-induced errors on high-confidence mentions.
Lorenzo Jaime Yu Flores, Cesare Spinoso-Di Piano, Jackie Chi Kit Cheung. Confident in a Confidence Score: Investigating the Sensitivity of the Calibration of Confidence Scores to Supervised Fine-Tuning in Generation Tasks
Uncertainty quantification (UQ) techniques measure confidence in language model outputs to support critical applications like hallucination detection and selective prediction. While prior work has developed various confidence metrics and demonstrated their calibration for classification tasks or using verbalized confidence, the robustness of probability-based and self-consistency-based UQ metrics for natural language generation (NLG) remains underexplored—particularly under model adaptation. Since practitioners routinely apply supervised fine-tuning (SFT) to adapt models to new tasks, a key question arises: do confidence metrics maintain their calibration when models are fine-tuned? We investigate this question across NLG tasks including translation, question answering, and mathematical reasoning. We find that calibration shifts substantially after SFT: across 216 configurations, it degrades in 112 cases and improves in 104, with confidence scores shifting due to factors beyond output quality, such as proximity to the training distribution. Degradation is therefore neither universal nor rare, and its direction cannot be anticipated from the pre-SFT model. Through a downstream task evaluation, we show that this miscalibration substantially reduces the practical utility of confidence scores for identifying correct answers. Our findings reveal that existing confidence metrics for NLG cannot be reliably deployed off-the-shelf after fine-tuning, highlighting the need for calibration-robust UQ methods under model adaptation.
Renato Okabayashi Miyaji, Matias Valdenegro-Toro. Rethinking Aleatoric Uncertainty: Partitioning Irreducible States for Large Language Model Agents
While Uncertainty Quantification (UQ) remains an active research field for Large Language Models (LLMs), the current literature lacks a rigorous understanding of what constitutes aleatoric uncertainty and its exact sources. Existing methodologies largely rely on the traditional dichotomy of aleatoric and epistemic uncertainty, which inadequately treats uncertainty as a static snapshot. To bridge this gap, this position paper challenges the static treatment of uncertainty. Supported by empirical evidence, we posit that aleatoric uncertainty in LLM agents is fundamentally dynamic and must be partitioned into strictly and temporarily irreducible states—differentiating intrinsic task variance, such as multiple valid answers, from resolvable noise, including input ambiguity, typographical errors, and contextual conflict. Through episodic empirical evaluations across multi-turn scenarios, we demonstrate how targeted agentic interventions can actively resolve temporarily irreducible noise, transitioning the interaction into an epistemic regime. When data uncertainty is strictly irreducible, model interventions cannot isolate a single correct distribution. While we acknowledge that the traditional static dichotomy remains a valuable framework for non-agentic settings, we urge the community to develop new decomposition paradigms that track uncertainty along the timeline of interaction. Embracing this dynamic perspective is essential to achieve robust uncertainty quantification for modern LLM agents.
Philip Müller, Nicholas Popovič, Michael Färber, Peter Steinbach. Semantic Calibration Prevails Where Token Confidence Fails: Benchmarking Long-Form Scientific QA
Reliable uncertainty quantification (UQ) is essential for safe deployment of large language models (LLMs) in scientific question answering, where long-form outputs exceed practical human verification at scale. We introduce the first large-scale benchmark for UQ calibration in long-form, reasoning-demanding scientific QA, evaluating four UQ methods on 685,000 responses across up to 20 LLMs and seven datasets, supported by an extensible open-source framework whose shared-generation design enables reproducible cross-method comparisons. Instruction tuning is shown to associate with systematic token probability polarization, collapsing confidence distributions and undermining the reliability of token-level uncertainty signals. Reasoning model families diverge: some reproduce this polarization while others actively mitigate it, a pattern that clusters by provider and suggests training pipeline design as a key differentiating factor. Verbalized and token-aggregation sequence-level methods fail systematically. Only semantic consistency, as measured by consistency of the final answer, yields well-calibrated outputs, providing the first large-scale evidence that semantic calibration persists in multi-step, dependency-rich reasoning settings.
Norbert Braunschweiler, Rama Doddipatla, Nigel Collier. Tokenization Risk as a Control Signal for Adaptive RAG
Retrieval-Augmented Generation (RAG) systems can degrade on inputs containing specialized, low-frequency vocabulary, where subword tokenization introduces fragmentation that disrupts both retrieval and generation. We introduce tokenization risk, a lightweight metric that predicts these failures. We show that (i) tokenization risk strongly correlates with degradation in retrieval and answer quality, and (ii) static mitigations (e.g., hybrid retrieval) provide only partial gains. To address this, we propose risk-based routing, which adaptively switches retrieval and generation strategies based on risk. Experiments on SpecialVocabBench, a risk-stratified benchmark of technical documents, demonstrate that this approach consistently improves retrieval accuracy and answer correctness with small system changes, establishing tokenization risk as an effective control signal for adaptive RAG.
Viju Sudhi, Regina Stodden, Philipp Cimiano. AttentionCD: Reducing Factuality Hallucinations in Large Language Models using Task-specific Head-based Contrastive Decoding
Despite the tremendous volume of data they are trained on, Large Language Models (LLMs) can generate plausibly sounding but non-factual responses. These factually hallucinated responses reduce users’ trust in such models and hinder their adoption in safety-critical applications. Among the many training-free hallucination mitigation strategies proposed, contrastive decoding strategies have received wide attention. Most of the work in this direction explores how contrasting probability distributions from different decoder layers can help in reducing factuality hallucinations. This work, in contrast, investigates whether masking of attention heads can provide reliable contrastive signals. In particular, we propose AttentionCD, a decoding strategy that masks task-specific attention heads to yield better contrastive signals to reduce factuality hallucinations in LLMs. Our approach shows competitive performance with the existing layer-based contrastive decoding strategies and offers an average performance improvement of over +11.1% over greedy decoding across different tasks and models.
Ieva Raminta Staliunaite, James Bishop, Andreas Vlachos. The Role of Ambiguity in Error Prediction via Uncertainty Quantification
The task of Error Prediction, namely predicting whether a model output is correct, is commonly tackled with Uncertainty Quantification (UQ). However, while uncertainty metrics capture when models lack knowledge or capacity to make a prediction, they also reflect aleatoric uncertainty, inherent in the input and context. We improve error prediction for Large Language Models (LLMs) by disentangling input ambiguity from the UQ signal. We experiment on Question Answering (QA) with six UQ metrics and show that UQ metrics are more predictive of errors on unambiguous instances than on questions with multiple plausible answers. We use Gated Experts and Selective Prediction to incorporate gold and predicted ambiguity labels into the error prediction pipeline. We find that ambiguity information improves error prediction across model families, training and evaluation paradigms, and datasets (including allegedly unambiguous ones), with the gains concentrated on the UQ metrics that most entangle aleatoric and epistemic uncertainty, gaining over 10 PRR points for individual metrics.
Seon Kim, Jeong-woo Lee, Tae Hoon Kong, Jongchan Kim. Uncertainty Meets Conformance: A Process Mining-based Multi-Objective Evaluation for Clinical Reasoning using LLMs
We present a framework for evaluating LLM-based clinical reasoning using conformance metrics from process mining. Four local LLMs were tested on EMR data from 16 vertigo patients across four prompting methods, with five repetitions per combination, yielding six uncertainty metrics. Reasoning traces were converted into event logs and analyzed with Heuristic Miner to derive external conformance to other patients’ reasoning patterns and internal conformance across repeated runs. Treating these as joint objectives, Pareto, Hypervolume, and TOPSIS analyses identified distinct evaluation scenarios, showing that conformance metrics can complement outcome-based uncertainty by revealing how LLMs reach their conclusions.
Tingxin Yang, Zefeng Wang, Mengyue Wang, Jinhe Bi, Xingcheng Zhou, Volker Tresp, Yunpu Ma. DeepLook: Uncertainty-Guided Branching with Lookahead Confidence for Efficient Reasoning
Test-time scaling improves large language model reasoning, but spends compute uniformly along a reasoning trace, even in stretches that are already stable. Motivated by the observation that reasoning failures show an early onset of uncertainty, we introduce DeepLook, a training-free decoding framework that monitors segment-level confidence, branches only at uncertainty bottlenecks, and ranks candidate continuations by Average Lookahead Confidence (ALC)—their mean confidence over fixed-horizon rollouts—before completing and voting over a pruned subset. On four competition mathematics benchmarks across four open models, DeepLook matches or exceeds the accuracy of DeepConf-low in 12 of 16 settings while generating 87.3\% fewer tokens, showing that a model’s own confidence is informative enough to decide where extra decoding compute is worth spending. On knowledge-intensive GPQA-Diamond the accuracy advantage disappears even though the token saving does not: coherent hallucinations also receive high ALC, so internal confidence localizes instability, not falsehood. Code will be released upon acceptance.
Katarina Pejcinovic, Marianne Casilio, Robert C Gale, Gerasimos Fergadiotis, Steven Bedrick. “But Isn’t It All Just Semantics?” The Importance of Evaluating Uncertainty, Reliability, and Random Effects
Picture naming tests, which involve showing images and asking patients to name them, are widely used clinical and research tools to evaluate a patient’s ability to retrieve and produce words. One essential aspect of such tests is categorizing the type of semantic similarity between the image and the patient’s response, as classifying errors is important for treatment planning and prognosis. These tests are tedious to score and prone to inter-rater variability, both of which could be improved with automation via LLMs. This task highlights the clinical potential of LLMs, but also the need for robust statistical testing of LLM performance. In this work, we show that traditional evaluation methods can be misleading for complex experimental designs. Using generalized linear mixed effects models and conformal prediction for uncertainty quantification, we identify sources of variation in model performance and estimate the uncertainty of model predictions. We show that seemingly large differences in model performance from traditional evaluations are misleading, driven instead by item-level and category difficulty. Additionally, uncertainty in predictions varies substantially between semantic categories. These results suggest that for NLP evaluations with nested, clustered, or uneven category distributions, uncertainty quantification can be necessary to avoid misleading conclusions for high-stakes clinical tasks.
Mayur Parvatikar. Constrained Decoding Can Reverse Which Inputs Look Uncertain
Systems that must return valid JSON or a well-formed function call can request the format in the prompt or enforce it with constrained decoding, where a grammar blocks any token that would break the format and the remaining probabilities are rescaled to sum to one. Such systems may also read the entropy of that rescaled distribution as a confidence signal. When the grammar blocks the tokens holding most of the model’s probability, what remains had received little, and rescaling it can make a concentrated distribution look diffuse. We show this in an adversarial setting built for the purpose, a safety-aligned model refusing a harmful request the grammar does not let it express, against benign instructions. Reading both entropies at the first generated token on Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct, $50$ prompts per class, the class ordering reverses, at AUROC $0.125$ to $0.175$ before the grammar and $0.613$ to $0.750$ after. Within a class the two readings are uncorrelated, so recalibration does not recover the pre-mask ordering. The two entropies measure different things, so systems should state which one they read, use the pre-mask one when the goal is the model’s own preference, and otherwise report how much probability the grammar let through. The reversal is confined to the first generated token and disappears once the prompt asks for the format, although the two readings then need not agree.
Raúl Vázquez, Aman Sinha, Chuyuan Li, Artem Shelmanov, Artem Vazhentsev, Claudio Savelli, Eduardo Calò, Emilio Raimond, Flavio Giobergia, Hengyu Luo, Jörg Tiedemann, Lorenzo Vaiani, Stella Frank, Vincent Segonne, Timothee Mickus. Overview of SHROOM-Visions 2026: A Shared Task on Hallucination Detection in Large Vision-Language Models
In 2026, we held the fourth iteration of the SHROOM Shared Task series: SHROOM-Visions (Shared-task on Hallucinations and Related Observable Overgeneration Mistakes in Vision language models), which is hosted at the UncertaiNLP Workshop co-located with EMNLP 2026. Following the success of the 2024 and 2025 tasks, this time we aim to tackle hallucinations through a model-agnostic detection task focused on large vision-language models. Building on the recently introduced SHEEP dataset, designed for long-term evaluation across model generations, the task invites participants to detect and classify fine-grained hallucination spans in image-conditioned text generation (VQA, image captioning, etc.). The evaluation uses a five-class taxonomy of hallucinations spanning four languages: Chinese, English, French, and Italian. The shared task generated strong interest in the NLP community worldwide, with 27 teams contributing 600+ system submissions. The best systems achieve average scores of 0.58 in character-level correlation, 0.46 in label-conditioned correlation, and 0.51 in intersection-over-union (IoU) across four languages, outperforming the baselines by 30–40 points.
Leon Eshuijs, Archie Chaudhury, Alan McBeth, Ethan Nguyen. But what is your honest answer? Aiding LLM-judges with honest alternatives using steering vectors
LLM-as-a-judge is widely used as a scalable substitute for human evaluation, yet judges are least reliable on subtle dishonesty such as sycophancy and manipulation, where a judge scoring a response in isolation lacks a reference point. We introduce Judge Using Safety-Steered Alternatives (JUSSA), a framework that reduces this evaluator uncertainty using the assessed model’s own internal representations. An honesty-promoting steering vector, optimized from a single training example, generates a contrastive alternative against which each response is judged. On a new manipulation dataset with human-validated response pairs, JUSSA improves detection for both GPT-4.1 (AUROC 0.893$\rightarrow$0.946) and Claude Haiku (0.859$\rightarrow$0.929) judges. The gains concentrate almost entirely on the pairs a judge would otherwise misrank and reverse once a judge has saturated a task, so contrastive evaluation helps when the judge retains headroom rather than as a function of model size. A layer-wise analysis further locates steering effectiveness in the middle layers, where model representations of honest and dishonest prompt processing diverge. Our work demonstrates that steering vectors can serve as tools for evaluation rather than for improving model outputs at inference, opening a new direction for white-box auditing.
Md Rayhanul Masud, Md Rizwan Parvez. TAME: Token Attribution and Masking for Emergent misalignment
Fine-tuning an aligned language model on narrow, flawed data can induce harmful behavior far outside the training domain, known as emergent misalignment (EM). Prior work has localized EM in model weights, activations, and training documents, but it remains unclear which training tokens carry the relevant fine-tuning signal. We introduce TAME (Token Attribution and Masking for Emergent misalignment), a three-stage framework: token attribution scores how strongly the fine-tuning update raises each response token’s likelihood, using forward passes through a released LoRA adapter; signal characterization finds patterns among high-attribution tokens; and causal validation tests them by attribution-guided loss masking. On released EM organisms and a 6,849-example medical-advice split, attribution is concentrated (top 5% of tokens hold 32% of the mass) and, in Llama, depleted for medical vocabulary but enriched for a register of unwarranted certainty, even after controlling for token rarity. Masking high-attribution tokens during fresh fine-tuning cuts EM 23× in Llama and 36× in Qwen, with the perplexity cost concentrated on the targeted register rather than on medical content; an equal random mask leaves EM unchanged. In Llama, the attribution pattern suggests that EM- relevant signal lies more in how confidently flawed content is expressed than in its domain vocabulary; the causal masking effect itself holds across both families.
Dries Wedda, Matias Valdenegro-Toro. Overlap-Based OOD Detection in Closed-Book Question Answering
Closed-book question answering requires models to answer using knowledge stored in their parameters, making uncertainty estimation important when that knowledge is limited. We study how total, aleatoric, and epistemic uncertainty respond to progressively weaker train-test overlap, from question paraphrase overlap to answer-only overlap to no overlap, and how well these estimates detect overlap-based out-of-distribution questions. Using T5-Large-SSM on WebQuestions and Open Natural Questions, we estimate token-level uncertainty with Deep Ensembles and Monte Carlo Dropout and sequence-level uncertainty with Deep Ensembles. Weaker overlap is associated with higher total and epistemic uncertainty, but aleatoric uncertainty also tends to increase. Epistemic uncertainty performs best for OOD detection on Open Natural Questions, but does not consistently outperform total or aleatoric uncertainty on WebQuestions. These findings show that train-test overlap is reflected in uncertainty estimates, but the response to weaker overlap is not consistently isolated to the epistemic component.
Yulin Chen, Shubhalaxmi Mukherjee, Pradeep K. Murukannaiah. Uncertainty-Aware Fact Checking with LLMs
Large Language Models (LLMs) offer a scalable approach to fact-checking, but their parametric knowledge remains unreliable for complex, specialized claims. Existing solutions unconditionally perform verification using Retrieval-Augmented Generation (RAG), which is computationally expensive and susceptible to performance-limiting noise. We propose an adaptive fact-checking pipeline that invokes RAG only when the LLM’s uncertainty surpasses a calibrated threshold. We introduce two UQ methods: (1) Adversarial Consistency Evaluation (ACE), a novel approach that detects model uncertainty by assessing its logical consistency, and (2) a trained ensemble that combines the strengths of several whitebox UQ methods and ACE. When evaluated on the SciFact and QuanTemp benchmarks, our adaptive pipeline preserves at least 88\% of the top-performing baseline’s accuracy. Simultaneously, our approach reduces retrieval frequency and token consumption by up to 55.3\% and 37.1\%, respectively, consistently outperforming decomposition-based RAG baselines at a fraction of the cost.
Eleni Partalidou, Tatiana Passali, Chrysoula Zerva, Grigorios Tsoumakas, Sophia Ananiadou. From Pixels to Tokens: Adaptive LLM Decoding and Reliability Analysis with Otsu Thresholding
This paper investigates the relationship between generation-time uncertainty and output reliability in probabilistic Large Language Model (LLM) decoding. Specifically, we assess the correlation between several factuality, quality and certainty metrics to infer existing patterns across state-of-the-art decoding strategies, diverse tasks and instruction-tuned models. We also propose a novel, image processing-inspired decoding approach, Otsu Thresholding, that performs adaptive, parameter-free threshold selection directly from the next-token predictive confidence distribution. The proposed Otsu-based variants achieve competitive quality and factuality across tasks and models, while exhibiting concentrated predictive distributions comparable to strong truncation-based baselines. Furthermore, we employ a selective filtering setup to analyze the impact of removing highly uncertain instances on the quality and factuality metrics, ultimately revealing that filtering based on averaged token-level uncertainty metrics can be practically useful for identifying less reliable generations, with filtering frequently improving the quality or factuality of the retained outputs.
Irina Proskurina, Mayank Kumar, Oyindolapo Olabisi Komolafe. Are You Sure You’re Sure? On the Impact of Instruction Tuning on Confidence and Lexical Diversity
Instruction-tuned language models achieve strong performance across a range of generation tasks but have recently been shown to exhibit verbalized overconfidence, which may manifest in less diverse supporting rationales for incorrect answers. However, whether such overconfidence is associated with rationale consistency remains an open question. In this paper, we study whether changes in the lexical diversity of generated answer rationales accompany changes in model confidence induced by instruction tuning. We evaluate three matched base and instruction-tuned models across question-answering benchmarks and find that instruction tuning consistently increases answer confidence, despite limited changes in predictive accuracy, while degrading likelihood-based calibration. Secondly, we observe a non-uniform effect of instruction tuning on rationale diversity: cross-rationale diversity consistently decreases, whereas surface-level lexical diversity varies in both direction and magnitude across models and benchmarks. Finally, we find that these differences persist after controlling for answer selection and rationale length, confirming that confidence and rationale diversity capture distinct effects of instruction tuning.
Aravind Mandiga, Guoming Li, Jin Lu. Paraphrase-Induced Retrieval Instability: Systematic Measurement and the Limits of Inference-Time Mitigation
Dense retrieval can return different evidence when a query is paraphrased, even when the meaning stays the same. This means that linguistic variation can create uncertainty in the evidence given to a retrieval-augmented generation (RAG) system. We study this problem across HotpotQA, FEVER, and SQuAD 2.0 using five dense encoders spanning 22M to 8B parameters and multiple model families. Paraphrase Jaccard distance remains 32 to 50 percent, showing that retrieval instability persists with stronger and alternative encoders. We then ask how much this problem can be reduced at inference time. Multi-source consensus over paraphrased queries and different retrievers improves retrieval coverage. Fixed source-type weights can help in harder settings. However, adapting those weights to each query using retrieval stability changes Coverage@20 by at most ±0.3 percentage points per setting. Retrieval stability is more useful as a label-free routing signal: on HotpotQA/MiniLM, routing only unstable queries through the stability-conditioned consensus configuration reduces retrieval calls by 47 percent for a 0.6 percentage-point Coverage@20 cost compared with always using that configuration. Overall, paraphrase-induced retrieval instability is a persistent source of uncertainty in retrieved evidence, and the query-adaptive stability component provides only a small marginal refinement.
Gaurav Negi, M.A. Waskow, John Philip McCrae, Omnia Zayed, Paul Buitelaar. Large Language Models as Automatic Annotators and Annotation Adjudicators for Fine-Grained Opinion Analysis
Fine-grained opinion analysis enables a detailed understanding of expressed sentiments and their corresponding targets. However, annotating domain-specific datasets requires substantial human effort and financial resources. To address this limitation, we investigate the feasibility of using large language models (LLMs) as automatic annotators for fine-grained opinion analysis. We use a declarative annotation pipeline to reduce prompt engineering when identifying fine-grained opinion spans, and compare rule-based voting against an LLM-based adjudicator for reconciling conflicting predictions. We trial the pipeline across different model sizes on the Aspect Sentiment Triplet Extraction (ASTE) and Aspect-Category-Opinion-Sentiment (ACOS) tasks. Our findings indicate a critical performance bifurcation: LLMs are reliable at the span level yet struggle to faithfully reproduce the relational structures connecting those spans. The better aggregation strategy depends on this complexity: voting suffices for triplets, while quadruples require an adjudicator that re-reads the source. Treating annotator disagreement as an informative signal rather than noise, we further show that agreement between LLM annotators is a calibrated predictor of correctness and can be used to delegate cases with high annotator disagreement to humans.
Kulin Nikita, Zhuravlev Viktor, Khairullin Artur, Sergey Muravyov, Ilya Makarov, Daniil Sukhorukov. BRAVE: Budget-Aware Reflective AutoPrompting for Efficient Evolutionary Prompt Optimization
Automatic prompt optimization is an uncertain sequential decision problem: an editing operator’s quality gain and context-dependent token cost are unknown before execution. We introduce BRAVE (Budget-Aware Reflective AutoPrompting for Efficient Evolutionary Prompt Optimization), a population-based optimizer that treats this operational uncertainty as a control signal. Its EVC Controller maintains per-operator Bayesian models of quality gain, token cost, and improvement probability, then selects among eleven generation and mutation operators by uncertainty-penalized predicted utility per token under a predicted-cost feasibility guard. An EMA return-on-investment mechanism revises this decision process from realized outcomes, while diversity management preserves alternative prompt hypotheses. On seven datasets, BRAVE outperforms EvoPrompt, PromptBreeder, and GEPA on 7/7 tasks; its advantage over the strongest external baseline ranges from 0.002 to 0.150 absolute score points. Under 20k, 40k, and 80k nominal budgets, BRAVE respectively achieves average quality 0.779, 0.807, and 0.812 versus GEPA’s 0.771, 0.775, and 0.781, while using 21.4k, 41.0k, and 72.4k tokens versus 46.1k, 65.2k, and 96.5k. BRAVE therefore formulates reliable prompt optimization as uncertainty-aware resource allocation: uncertainty about an edit’s value and cost directly changes the action selected.
Ananto Nayan Bala. Does Quantization Change What Agreement Means? A Paired Audit of Homogeneous LLM Panels
Agreement among repeated large language model (LLM) responses is often used as a confidence signal. We test whether that signal changes after quantization by comparing five-response panels from the same checkpoints at FP16 and 4-bit NF4. We evaluate two open models on 300 TruthfulQA multiple-choice questions and 300 GSM8K problems, matching prompts, sampling settings, and seeds. Under strict parsing, quantization reduces GSM8K panel accuracy by 10.5 percentage points for Qwen2.5-3B and 6.3 for Phi-3.5-mini. The probability that all five responses are wrong rises by 9.0 and 5.0 points. Agreement also becomes less useful for selecting reliable answers, but all-wrong panels do not consistently become more likely to give one identical answer. Relaxed parsing preserves the selective-reliability result for both models, although Phi’s panel-accuracy and all-wrong differences become inconclusive. On TruthfulQA, Phi degrades while Qwen is largely stable. Quantization can therefore reduce panel reliability without consistently pushing all members toward the same wrong answer.
Toshiki Kuramoto, Ryohei Kamei, Jun Suzuki. What Counts as Hallucination? A Scoping Review of Tasks, Symptoms, and Convergence Expectations in LLM Research
Hallucination in large language models (LLMs) and related generative models has become a central issue for reliability and practical deployment. Yet prior work does not consistently define what is treated as hallucination or the level of expectation under which model outputs are evaluated. This study conducts a scoping review of 244 papers published between 2018 and 2025, analyzing definitions of hallucination, output-side symptoms, and Convergence Expectation (CE). The results show that approximately 90% of the papers provide some form of operational definition, but the symptoms targeted by these operationalizations vary substantially by task. “Addition” appears across many tasks as a common symptom, whereas “Transformation” and “Deviation” strongly depend on task structure and input constraints. Analyses by CE band further show that, even within the same task, the symptoms problematized as hallucination vary according to the strictness of expected output convergence. These findings indicate that hallucination is not a fixed failure inherent in model outputs, but a phenomenon shaped by task, input constraints, evaluation criteria, and the expectations of humans or evaluators. This study reframes hallucination as expectation mismatch and argues that definitions, symptom scope, and expectation levels must be made explicit when comparing research findings and designing evaluations.
Emil Kalbaliyev, Kairit Sirts. Refer, Infer, or Abstain: Investigating Latent Skill Selection of Large Language Models under Contextual Question Answering Uncertainty
Contextual question answering (CQA) involves uncertainty regarding the use of available knowledge in forming an answer. In CQA, large language models (LLMs) should decide whether to Refer to explicitly stated information, Infer from contextual evidence using commonsense knowledge, or Abstain when an answer cannot be supported. We investigate whether LLMs implicitly select among these skills under CQA uncertainty. Combining prompting analysis with layer-wise probing, we examine latent skill selection across instruction-tuned LLMs. We find that the models distinguish Refer from Infer more easily than Infer from Abstain. In particular, erroneous abstentions disproportionately occur on inference-based questions, and the Infer–Abstain distinction is the least separable in model representations. We further find that models represent contextual knowledge gaps more strongly than contextual knowledge sufficiency. These results characterize the emergence and limitations of latent skill selection in LLMs under CQA uncertainty.
Hayden Moore, Abhinav Verma, Anshuman Dutt. Tracing Recurrent and Emergent Hallucination Patterns in Decomposed Text-to-SQL Generation
Large language models can achieve strong Text-to-SQL benchmark performance while still producing schema and query-logic errors. Most evaluations characterize such failures only in the final generated query, making it difficult to determine whether a related error was also observed in an earlier, simpler step or appears only at the final stage. We present a drill-down evaluation framework that decomposes Text-to-SQL queries into progressively more complex executable sub-queries and corresponding natural-language questions, allowing rule-based hallucination categories to be examined across a generation path. We distinguish recurrent hallucinations, in which the same rule-based category is observed in an earlier step and again in the final query, from emergent hallucinations, in which the category is observed at the final step without an earlier detection in the same path. Across five LLMs evaluated on an expanded BIRD-mini benchmark, several categories show appreciable recurrence probabilities, while clause abuse shows comparatively higher conditional emergence. Under the conversational-history condition, recurrence probabilities are generally higher and emergence probabilities are generally lower. These observations indicate that the two conditional measures do not move uniformly in the same direction under additional context and motivate trajectory-level analysis in addition to final-output evaluation. Code and evaluation artifacts are available in the project repository: https://github.com/HaydenMM/bird-mini-drilldown
Claudio Moisés Valiense de Andrade, Guilherme Fonseca, Washington Cunha, Celso França, Adriana Silvina Pagano, Leonardo Chaves Dutra da Rocha, Marcos André Gonçalves. Beyond Global Calibration: Analyzing Class-Specific Miscalibration and High-Confidence Errors in Text Classification
Confidence scores are increasingly used to assess uncertainty in text classification, yet aggregate calibration metrics may hide localized reliability failures. We investigate this issue across seven sentiment and topic classification datasets by comparing global and class-specific calibration and analyzing high-confidence errors. Our results show that low global calibration error can coexist with substantial class-specific variation, particularly in multi-class settings. We also find that high-confidence errors are not randomly distributed: they concentrate in specific true–predicted class confusions, with some incorrect predictions receiving confidence scores close to 1.0. These findings show that globally calibrated models can still exhibit systematic uncertainty failures, motivating class-aware and error-level analyses when confidence is used as an indicator of prediction reliability.
Soham Tripathy. Confidence Drift During Task-Specific Language Model Fine-Tuning
Task-specific fine-tuning can change not only what a language model predicts, but also how confident it is in those predictions. We study confidence and accuracy drifts throughout LoRA fine-tuning of seven decoder-only language models spanning multiple families and scales. Models are fine-tuned on AG News and evaluated at intermediate checkpoints on the source task and five untouched cross-task benchmarks. We find that cross-task accuracy generally remains stable or degrades, while confidence increases on both correct and incorrect predictions. Across many settings, confidence follows a non-monotonic trajectory: after an early transient and a relatively stable middle stage, it rises again late in training, coinciding with increasing train–validation NLL divergence. Our results show that accuracy-only evaluation can miss confidence drifts introduced by task-specific fine-tuning.
Maciej Skorski. Moral Entropy: Auditing Bias and Uncertainty in Moral Judgment
Most work in computational ethics treats annotator disagreement on moral content as noise to be voted away, collapsed into majority vote or the more permissive any-annotator rule which assigns a label as soon as any annotator flags it.
We argue this uncertainty should instead be modeled and learned from.
We introduce Moral Entropy, a Bayesian framework that estimates a full posterior over the true label and decomposes its entropy into aleatoric uncertainty (irreducible disagreement about the moral content) and epistemic uncertainty (from insufficient or noisy annotation). This lets any heuristic consensus rule be audited against a calibrated ground truth via entropy methods such as cross-entropy/KL, Brier score, and expected calibration error.
Running this model on over 100,000 posts from three corpora spanning fifteen discourse domains, we audit the standard aggregation rules against the resulting posterior and reveal systematic differences in moral labels. Across MFTC and MFRC, the any-annotator rule yields mean false-positive rates of $10.2$–$19.9\%$ and mean false-negative rates of $38.9$–$54.8\%$ relative to the Bayesian consensus, while majority-vote and two-vote rules miss $63$–$83\%$ of reference-positive labels.
Kabir Murjani, Parth Vyas, Akshita Abrol, Rajesh Gupta, Sudeep Tanwar, Zhengkui Wang. The Variance of a Judge: Zero-Variance Rewards by Deterministic Regression
Large Language Model (LLM) judges score responses by generating rating tokens, so their rewards carry uncertainty from the scoring mechanism itself, not only from genuine disagreement about quality. Sampling variance, prompt phrasing, and self-preference bias inject epistemic, generation-induced noise. Neighbour-Oriented Reward Anchoring (NORA) replaces generative scoring with retrieval-augmented regression: a small MLP, the Optimised Response Latent Inference (ORLI) judge, scores each query-response pair over a frozen embedding manifold anchored to human-rated neighbours. Because no language model generates the score, the reward is deterministic and its generation-induced variance is zero by construction, leaving a residual error (MAE=0.736 on the 1-5 scale) with no run-to-run component. This relocates the error rather than removing it: a per-tier breakdown shows the residual is dominated by systematic bias, the scorer compressing predictions toward the interior of the scale, rather than by aleatoric annotation noise. ORLI is still a learned reward, so NORA curbs over-optimisation with a retrieval-guided curriculum and a contrastive penalty against known-poor examples. Improvements are asymmetric and largest where quality is lowest: a 0.5B actor on Recipe4U raises 1-2 star responses by up to +0.77 ORLI points while leaving 5-star outputs essentially unchanged, the pattern holds on HelpSteer2, and a blinded human evaluation reproduces it. Ablations identify the curriculum as the primary contributor.
Binesh Arakkal Remesh, Thomas Gerald, Sahar Ghannay, Eva Feillet, Camille Guinaudeau. Detection of parametric or contextual knowledge: are the schemas similar across modalities?
In LLMs, the source of information used to produce an answer is necessary to explain the generation process.This work considers the case where the answer can be extracted from the question’s context ( contextual knowledge) or from the model weights (parametric knowledge). While methods have been developed to uncover the grounding sources of LLMs in question answering, these approaches mainly rely on the text modality, whereas most tasks involve multiple modalities. This study proposes to uncover LLMs’ internal signals to determine whether they use parametric or contextual knowledge in a bimodal setting. To this end, we create a counterfactual dataset to evaluate whether the model relies on parametric or contextual knowledge on text and/or image-based questions. We leverage probing methods based on the internal attention and the different latent representations to determine the model’s source of information. We evaluate the performances of the probing methods, focusing on the different modalities. We show that models share common signals across the corpus and modalities. This study opens the way to develop controllable methods by uncovering common signals in LLMs to determine the knowledge used in generation.
Vani Kanjirangat, Oscar William Lithgow-Serrano, Alessandro Antonucci. Credal Ensembles for Uncertainty Quantification in LLMs
We consider the problem of uncertainty quantification in LLMs from a black-box perspective. To effectively disentangle aleatoric and epistemic uncertainty, we use an ensemble method that produces responses paired with so-called \emph{credal} sets of distributions. We combine different ensembling techniques, based on temperature variation, multiple models, paraphrasing, and permutations, with semantic-similarity and verbalized elicitation methods, resulting in new, computationally cheap measures of uncertainty. A benchmark on closed-ended (BoolQ and CommonsenseQA) and open-ended (AmbigQA) questions shows good correlation with predictive accuracy. Notably, when considering ambiguous tasks, our analysis reveals a tendency of LLMs to focus on a single interpretation, an important flaw in current LLM-based solutions.
Yuxi Xia, Dennis Ulmer, Terra Blevins, Yihong Liu, Hinrich Schuetze, Benjamin Roth. Calibration Is Not Enough: Evaluating Confidence Estimation Under Language Variations
Confidence estimation (CE) indicates how reliable the answers of large language models are and impacts user trust and decision-making. Existing evaluations mainly concern the alignment between confidence and correctness, but ignore the variability of language: confidence estimates should remain consistent under semantically equivalent prompts or answer variations, while changing when answer meaning differs, as this may indicate a change in correctness. Therefore, we introduce a novel evaluation framework based on three complementary properties: robustness to prompt perturbations, stability across semantically equivalent answers, and sensitivity to semantically different answers. We show that these metrics are largely independent of existing CE metrics, and that common CE methods often fail on them: while most methods achieve high robustness and stability, they struggle to distinguish semantically different answers, potentially because they do not effectively leverage generation-side information. Overall, our framework exposes overlooked limitations of current CE evaluations and provides guidance for selecting confidence estimators for real-world applications.
Dayeon Jang, Yuseok Bae, Kimin Yun. Explicit State Gating and Implicit Refusal for Unreadable Scene Text
Scene text recognition systems must decide what to do when an image contains no reliably readable text. We compare two abstention contracts for a vision-language scene-text recognizer. Under implicit refusal, the model signals unreadability inside its own transcription output. Under explicit state gating, a separate five-way state classifier decides whether a transcription is emitted at all. Across four training variants sharing the same backbone, direct OCR yields similar accuracy on readable text; the two contracts differ mainly in how they handle inputs with no readable text. Gating reduces false transcription (emitting text where none is readable) from 45.9% to 1.7%, at the cost of falsely abstaining on 3.8% of readable inputs. Which contract is preferable, however, cannot be determined without first specifying how much a correct abstention is worth. When rejected inputs are excluded from scoring, all gated configurations show similar risk; when a correct abstention is credited as a useful outcome, gating lowers risk on the diagnostic set. We present this preliminary analysis not as a context-free model ranking, but as a framework for choosing an abstention contract at deployment time. These results motivate future systems that use the predicted state to select a recovery action before abstaining.
Sravan Karthick T, Pranav Darshan, Pranav A, Minal Moharir, Ivan P. Yamshchikov. The Temporal Tug-of-War: Visualizing and Detecting RAG Conflicts in Diffusion Models via Trajectory Variance
Retrieval-Augmented Generation (RAG) introduces a specific failure mode in discrete diffusion language models: when retrieved context contradicts parametric knowledge, the iterative denoising process becomes a visible battleground between competing knowledge sources. We identify temporal semantic divergence as an observable for detecting these conflicts and introduce the Trajectory Variance Score (TVS), a simple and interpretable measure of this divergence. TVS computes the mean pairwise cosine distance of answer embeddings across independent stochastic denoising trajectories, capturing the temporal tug of war between parametric and contextual attractors. Requiring as few as two parallel inference runs, TVS is computationally lightweight. Across four diverse datasets (Synthetic, SciQ, PopQA, and CounterFact), a simple Logistic Regression classifier using TVS achieves $70.10\%$ accuracy and $0.7647$ AUROC on LLaDA. On Dream 7B, increasing the number of trajectories from two to five improves accuracy from $63.91\%$ to $69.62\%$. More complex sequential models provide only marginal improvements over the linear classifier. Evaluation across LLaDA and Dream 7B demonstrates that conflict-induced trajectory dynamics and their key properties transfer across distinct diffusion architectures.
Samuel Forster, Inmaculada Arnedillo-Sánchez. FAST-Eval and FAST-LoRA: Uncertainty Evaluation Framework and Low-Rank Adaptation for Finding Attributes with Supporting Text
Many domains require reasoning over unstructured text to identify attributes of interest that are not stated in a standardised form. Useful judgements on the presence of these attributes should also express calibrated confidence statements and cite supporting text as evidence. We define this task as Finding Attributes with Supporting Text (FAST). We introduce FAST-Eval, an uncertainty evaluation framework for confidence calibration, attribute identification and evidential grounding in FAST, together with a dataset of 1,000 human-annotated (document, attribute) pairs validated against an independent recruitment expert (Spearman ρs = 0.853, binary κ = 0.715). Taking employability skills in CVs as a case study, we use FAST-Eval to test whether FAST-LoRA, a Low-Rank Adaptation (LoRA) adapter applied to Qwen3-8B, improves performance in these three dimensions. Trained on 500 pairs, FAST-LoRA outperforms its untuned Qwen3-8B counterpart and GPT-5 on unseen skill identification (F1 0.877 vs 0.736 and 0.766) and calibration, while matching GPT-5 on citation fidelity. Both untuned models assign the highest confidence level to about half their judgements; FAST-LoRA instead distributes confidence across the scale and makes fewer unsupported positive predictions. These results suggest that modest domain adaptation is a more effective strategy than scaling model parameters for FAST.
Laura Braun, Matthias Aßenmacher, Martin Hohenadler, Goeran Kauermann. Two Lawyers, Three Opinions: Can Output-Based LLM Uncertainty Find the Cases Where Experts Disagree?
Legal natural language processing (NLP) tasks often lack a single objective ground truth, as reflected in label variation across experts’ annotations. We test whether two uncertainty signals based on outputs of large language models (LLMs), sampling entropy and verbalized probability, can identify posts where legal experts disagree. On 1,564 German social media posts annotated by three legal experts for punishability under the German Criminal Code, we evaluate four open-weight LLMs, with reasoning either enabled or disabled, and observe a drop in F1 score on contested posts. Sampling distributions collapse to near-deterministic judgments in most conditions, while verbalized probabilities track expert disagreement in aggregate and support deferral of model errors. Neither signal, however, reliably identifies individual contested posts: both detect the model’s own errors (AUROC up to 0.83) better than expert disagreement (up to 0.76), and which signal is more informative depends strongly on model and activation of reasoning mode.
Thomas Reiter, Christoph Kern, Fedor Miasnikov, Sofiia Nikolenko, Rob Chew, Stephanie Eckman, Frauke Kreuter. Reliable but Design-Sensitive: Instrument Uncertainty in LLM Annotation
Large language models (LLMs) can give reliable labels under one setup yet change those labels when researchers make other reasonable design choices. We tested seven LLMs, 12 task designs, three independent runs, and 3,000 tweets labeled for offensive language and hate speech. Repeating the same model and task design produced high agreement (median Fleiss’ $\kappa = 0.91$). Agreement fell when we changed the task design for the same tweets (median Cohen’s $\kappa = 0.76$). Task design and model choice increased the variance of estimated prevalence by factors of 76.7 for offensive language and 110.6 for hate speech compared with sampling variance alone. Variation across LLM task designs reached 560–572 basis points, compared with 270–331 basis points across five human instrument versions. Confidence scores did not solve this problem. They tracked repeated model outputs more closely than agreement with human labels, and grouping six tweets in one prompt lowered mean offensive-language confidence by 660 basis points. We call the variation caused by task design and model choice \textit{instrument uncertainty}. Researchers can measure it only by comparing reasonable task designs. Repeating one setup or relying on confidence scores cannot replace that test.
Hunter Tuerlings, Matias Valdenegro-Toro. Verbalized Uncertainty in Coding LLMs: A Contamination-Aware Study of Confidence Calibration
Coding large language models (LLMs) can produce plausible solutions that fail hidden test cases, making it difficult to judge when to trust their output. This paper presents a contamination-aware comparison of established methods for eliciting verbalized confidence in generated code. Six instruction-tuned, coding-specialized LLMs from three model families are evaluated on 365 problems from a shared contamination-aware \textsc{LiveCodeBench} window. For each frozen code generation, confidence is elicited separately under three strategies, Direct numeric scoring, Chain-of-Thought reasoning followed by a score, and Linguistic confidence labels. Under strict scoring, all 18 model-strategy combinations have negative Net Calibration Error, showing a consistent tendency toward overconfidence. Linguistic elicitation achieves the lowest Expected Calibration Error for four of six models, but several favourable values result from collapsed confidence distributions. Eight of the 18 combinations are collapsed and five are unstable across repeated runs, including three Chain-of-Thought combinations. When outputs are ranked by confidence and the least confident are reviewed first, Direct elicitation gives the largest improvement over random selection in four of six models. Verbalized confidence cannot be reduced to a single trustworthy pass probability; calibration, ranking ability, distributional collapse, and stability must each be checked separately.
Non-archival
Yunxiang Li, Xixin Wu, Helen M. Meng. How Many Pixels Is a Digit Worth? Place-Aware Coordinate Entropy for GUI Agent Confidence Estimation
GUI agents predict click coordinates as digit-token sequences, but standard text-LLM confidence estimation methods rank correct clicks from wrong ones only weakly. GUI-specific alternatives use $K$ samples or new supervision, but still leave room for improvement. We trace part of this to \emph{place-value asymmetry}: bounding- box correctness often makes higher-place digits more important than lower-place digits, so uniform aggregation weakens the signal that determines correctness. The fix is to weight each digit’s Shannon entropy by its place value. We call this \textbf{Place-Aware Coordinate Entropy} (PACE). Across fixed-scale agents on ScreenSpot-Pro and ScreenSpot-v$2$, PACE wins both AUROC and selective accuracy on all primary comparisons in a single forward pass, matching or outperforming $K$-sample baselines at a fraction of the cost. PACE provides a per-click confidence estimate that turns coordinate-token internals into a practical confidence signal for GUI agent deployment.
Athina Kyriakou, Dennis Ulmer, Ivan Titov. Shared Doubt: Zero-shot Cross-Lingual Confidence Estimation for Language Models
Confidence estimation (CE), i.e. quantifying the reliability of a model’s prediction, has attracted great interest in the context of large language models (LLMs). However, most studies focus on English, ignoring the multilingual reality of LLM usage, while many CE methods degrade or require retraining across languages. To address this gap, we investigate whether multilingual LLMs encode shared, language-transferable confidence features. We use a lightweight linear probe that predicts answer correctness directly from intermediate representations. Trained monolingually, the probe generalizes zero-shot to \emph{unseen}, typologically diverse languages without target-language supervision. Learned layer weights and multiple ablations reveal that confidence features concentrate in middle layers across languages, suggesting a shared confidence subspace. While zero-shot cross-lingual performance depends on similarity to the source language, the probe provides a strong baseline without any retraining and compares favorably to other popular confidence estimation methods.
Timothee Mickus, Claudio Savelli, Eduardo Calò, Emilio Raimond, Stella Frank, Hengyu Luo, Flavio Giobergia, Vincent Segonne, Chuyuan Li, Aman Sinha, Lorenzo Vaiani, Jörg Tiedemann, Raúl Vázquez. Can Humans Dream of Electric Sheep? Human-Written Samples for Fine-Grained Vision-and-Language Hallucination Benchmarking
In an age of rapid model turnover, how do we make hallucination evaluation more perennial? We explore whether human-written hallucination samples could take the place of model-generated hallucinations, in order to make benchmarking detection independent of particular models. To this end, we construct SHEEP, a dataset of 1,600 human-written samples, spanning four languages (Chinese, English, French, Italian), and 18,400 samples from five vision-and-language models, all annotated for hallucinations using a fine-grained span-level labeling scheme. We find that human-written samples result in higher agreement and allow greater control of dataset contents, while remaining distributionally similar to samples derived from vision-and-language samples and providing a reasonable portrayal of detection capabilities — suggesting that human data is a viable substitute for model-based hallucination benchmarks.
Arya Hariharan, Ananth K. Suresh. CANDOR: Calibrated Answering across Knowledge Region-Boundaries
Knowledge boundaries of large language models (LLMs) define the limits of what models reliably know. Despite appearing well-calibrated, LLMs frequently exhibit overconfident failures near these boundaries, like on rare entities, outdated facts, subtly incorrect premises, or mixed-familiarity inputs. Existing benchmarks do not systematically probe how confidence and correctness interact under such conditions. To address this, we introduce CANDOR (Calibrated Answering across Knowledge Region-Boundaries), a benchmark jointly measuring factual correctness and expressed confidence across four boundary conditions, namely long-tail entity questions, non-existent entity probes, mixed-fact questions, and near-true distractors, to test LLM robustness and reliability. We formalize knowledge boundary violations along five dimensions and evaluate eight models using a hybrid confidence elicitation protocol requiring no model internals access. Our results show that no model handles boundary conditions reliably. We also show that asking the model to reason before answering causes only marginal gains, showing that calibration failures at knowledge boundaries are not reasoning failures and resist inference-time correction. CANDOR is publicly available with full evaluation code (https://github.com/Arya-Hari/CANDOR).
Joe Stacey, Hadas Orgad, Kentaro Inui, Benjamin Heinzerling, Nafise Sadat Moosavi. A Robust Evaluation of Probe Robustness: Lessons for Reliable OOD Uncertainty Quantification
Recent work has shown that the hidden states of large language models contain signals useful for uncertainty estimation, motivating a growing interest in efficient probe-based approaches. Yet it remains unclear how robust existing methods are, with prior work reporting conflicting conclusions under substantially different evaluation settings. We address this by introducing ProbeDrift, a systematic evaluation framework for supervised uncertainty probes covering a wide range of OOD settings across models, tasks, and distributional shifts. Using ProbeDrift, we train over 2,000 probes to disentangle the effect of key design choices, showing poor robustness beyond near-OOD settings. We find that robustness is driven by design decisions that have a largely invisible effect in-distribution, including the choice of feature type, aggregation strategy, and training signal. We argue that robust uncertainty estimation requires robust evaluation. To support this, we release ProbeDrift as a lightweight Python library that contains the train and test splits underpinning our extensive evaluation. Finally, we show how the insights from ProbeDrift can directly lead to more robust methods through a simple Hybrid Back-Off (HBO) strategy.
Yuelyu Ji, Wuwei Lan, Patrick Ng. MM-RAGChecker: Validated Claim-Level Diagnosis for Visual Retrieval-Augmented Generation
Evaluating visual retrieval-augmented generation (RAG) requires more than answer-level accuracy. A system may produce a correct short answer while grounding its long-form justification in the wrong modality, ignoring retrieved evidence, or hallucinating unsupported claims. We present \textbf{MM-RAGChecker}, a multimodal claim-level diagnostic framework for visual RAG. MM-RAGChecker decomposes long-form answers into atomic claims, verifies each claim against retrieved text and image evidence, and aggregates these judgments into interpretable metrics for hallucination, claim recall, context precision, evidence utilization, and cross-modal grounding balance.
To validate the framework, we evaluate both its individual components and its end-to-end diagnostic reliability on model-generated answers. We further deploy MM-RAGChecker on \textbf{MRAG-Suite}, a cross-domain visual RAG testbed spanning web photos, charts, scanned documents, slides, and scholarly PDFs under controlled ambiguity and distractor settings. Across evaluated vision-language models (\textbf{VLMs}), we find that (1) domains with similar EM can differ sharply in grounding quality, (2) retrieved text is often underused relative to images, and (3) ambiguity and distractors substantially worsen both answer accuracy and claim-level faithfulness. These results show that MM-RAGChecker surfaces failure modes that remain invisible under conventional answer-level metrics alone.
Ti Tze Hong, Saffron Salmah Yen Lim, Maarten Sap. Fortifying Hallucination Detection to Out-of-Domain Data
Hallucinations remain one of the major barriers to the reliable deployment of Large Language Models (LLMs). Recent works have explored hallucination detection using both supervised classification based approaches and unsupervised metric based approaches, with the latter remaining popular since they do not require labeled data. In this work, we study the performance and generalizability of these approaches through a comprehensive empirical analyses across 11 datasets and 10 models. We first establish that supervised methods achieve stronger in-domain performance but lag behind unsupervised methods in the out-of-domain setting. Then to narrow this generalization gap, we introduce a simple approach that improves the generalizability of supervised hallucination detectors by training on a curated, multi-domain data mix, which can complement the subsequent addition of task-specific data. In our experiments on hallucination detection on 697K QA samples from 11 open source QA datasets, we show that incorporating this general training allows supervised methods to surpass unsupervised metric based methods by an average of +7.25% on out of domain data, without addition of any task-specific data. We also analyze scaling behaviors and estimate how much task-specific data is required to achieve reliable performance, finding that models augmented with general data require up to 40.3% less task-specific data to achieve close to optimal performance. Together, our findings highlight a simple path toward fortifying supervised hallucination detectors against domain shift.
Timo Saala, Lucie Flek. When Measures Become Targets: Construct Validity Failures in Uncertainty-Based Hallucination Detection
Uncertainty-based LLM hallucination detectors operationalize factual incorrectness, a human-interpretable concept, through measurable proxies: distributional confidence, semantic consistency under sampling, or embedding geometry. We show that each operationalization introduces a structural gap between the proxy and the human concept, exposable and quantifiable through targeted perturbations designed to satisfy the proxy while violating the concept, or vice versa. We study three uncertainty-based hallucination detectors - First Output Token Entropy, Semantic Entropy, and EigenScore - using diagnostic perturbations targeting each detector’s structural assumptions. We introduce a diagnostic threat model covering false-negative and false-positive construct violations, and design detector-specific perturbations that induce proxy-concept divergence in each method. Across three question-answering datasets and three LLMs, all evaluated detectors exhibit substantial proxy-concept divergence under optimization pressure, with hallucinated outputs going undetected or correct outputs falsely flagged. Human evaluation confirms that metric verdicts and human judgements systematically and consistently disagree, establishing genuine construct validity failures and motivating robustness-oriented evaluation.
Shreya Rajpal, Tanawan Premsri, Parisa Kordjamshidi. Spatial Reasoning via Modality Switching Between Language and Symbolic Representations
Human reasoning is inherently multimodal: when problems become difficult, we rarely think in words alone. We often externalize our reasoning by sketching diagrams or drawing grids to understand the underlying conceptual structure and avoid mistakes. Building on this premise, our research investigates: (a) whether grounding multi-hop textual-spatial stories into geometry-aware modalities, such as layouts or grids, improves reasoning compared to natural language-based inference; and (b) whether a model can decide when to rely on natural language reasoning and when to switch to a structured modality. We address these questions by introducing a switching metric based on trustworthiness and complexity signals, which estimates when grounding a spatial story into a structure is likely to improve performance. This takes a first step toward principled modality selection in large language model (LLM) reasoning. Across our settings, switching from natural language-based reasoning to a grid-based representation improves LLM performance by up to 42\%, highlighting the importance of modality choice in shaping reasoning outcomes.
Neil Dixit. Drift-Aware Stability of Longitudinal NLP Measurements
Social NLP systems are often used not only to classify texts, but also to measure aggregate social phenomena such as toxicity, misinformation, sentiment, stance, and political framing. In longitudinal settings, researchers may interpret changes in average model scores as evidence of social change. However, text distributions shift over time as topics, events, entities, communities, and linguistic conventions change. We formalize longitudinal NLP measurement as estimating $\mu_t(f)=E_{X\sim P_t}[f(X)],$ where $P_t$ is the text distribution at time $t$ and $f$ is a measurement model. We show that empirical changes in $\hat \mu_t(f)$ are controlled by distributional drift and finite-sample uncertainty, yielding a scale against which raw trends can be interpreted. This motivates a Trend Stability Ratio comparing observed measurement changes with estimated drift and sampling noise. Across two time-indexed social-media settings, TSR confirms a dominant topic shift in one dataset while substantially re-ranking candidate trend changes in another. These results suggest that longitudinal NLP measurements should be interpreted alongside drift-aware stability diagnostics, rather than from raw changes in aggregate model scores alone.
Kun Efimov-Zhang, Yifei Song, Claire Gardent. XQDT: eXplainable and Quantitative Data-Text Alignment Metric with Feedback Signals
Evaluating data-text alignment remains challenging: existing metrics often provide limited explanations for the scores, while prompt-based LLM-as-Judge methods can be expensive and unreliable. We present an end-to-end explainable evaluation metric that fine-tunes a language model to identify omitted, extra, incorrect, and correct data units in a data-text pair. These local judgements are aggregated into precision, recall, and F1 scores, providing both fine-grained diagnostic feedback and an interpretable measure of alignment quality. Across benchmarks, our fine-tuned models outperform LLM-as-Judge methods in error prediction and surpass existing metrics in precision, recall, and F1, while maintaining strong correlation with human judgements. Beyond evaluation, our verifier outputs also provide useful feedback signals for downstream correction and refinement, supporting alignment-oriented improvement of data-to-text and text-to-data. This paper has been accepted to the EMNLP 2026 Main Conference.
Tianyi Ma, Parisa Kordjamshidi. CLAMP: Constrained Decoding for Vision-Language Embodied Planning
Embodied planning increasingly relies on vision-language models (VLMs) to translate instructions and visual observations into executable action sequences. However, fluent plans are not always executable. A VLM may refer to objects that are not visually observed, select actions whose required affordances are unavailable, or violate syntax and action constraints. We introduce CLAMP, a multimodal constraint-grounding framework that turns scene evidence into decoding-time constraints for a frozen VLM planner. CLAMP uses the initial observation to restrict object references to those supported by the scene, while a provided symbolic action model specifies state transitions and goals. During decoding, hard masks eliminate invalid next-token candidates, while a Hidden Markov Model (HMM)-based world-state lookahead module reweights the probabilities of the remaining feasible candidates based on action preconditions and goal reachability. This allows the planner to retain the VLM’s language prior while preventing visually unsupported, unsafe, or infeasible candidates from entering the plan. For unseen tasks and environments, CLAMP adapts the HMM at test time using label-free continuations sampled from the frozen VLM. Experiments on VLABench, SafeAgentBench, and TaPA show that scene-grounded constraints improve object grounding and safety, while most remaining failures stem from perception errors or misaligned constraint specifications.
Debarpan Bhattacharya, Malay Phadke, Sriram Ganapathy. BiG-SURE - Bipartite Graph Spectral energy for Uncertainty and Reliability Estimation of LLMs
Reliable uncertainty estimation is a crucial requirement for deploying large language models (LLMs) and vision-language models (VLMs) in safety-critical settings, especially when the model parameters are not accessible (black-box). We propose BiG-SURE, an uncertainty estimator based on cross-temperature semantic agreement. The method samples low-temperature responses as stable semantic anchors and high-temperature responses as probes under meaning-preserving input transformations. It then constructs an anchor–probe Bipartite Graph (BiG) using NLI-based entailment scores and defines confidence through the normalized squared spectral energy of this matrix, with uncertainty given by its complement. This measure, termed Spectral energy for Uncertainty and Reliability Estimation (SURE) measures whether high-temperature probes remain semantically aligned with the model’s stable low-temperature belief or not. We evaluate BiG-SURE on text QA, multilingual QA, and multimodal QA tasks across multiple model families. In these experiments, BiG-SURE improves average abstention AUROC over prior black-box uncertainty estimators, while remaining simple, unsupervised, and applicable to black-box model settings.
Eitan Wagner, Omri Abend. Words Can’t Describe: On the Limitations of Verbalized Confidence
How should language models report uncertainty over a set of possible answers? A common approach uses the token-level logprobs, which score each option independently but conflate linguistic output predictions with factual confidence. An alternative—described as verbalized confidence or second-order prediction—prompts the model to output confidence or probabilities as text, making uncertainty reporting part of the task itself. Although theoretically sound, verbalizing a distribution faces practical challenges by forcing the model to format and serialize an inherently unordered structure. We identify three costs of this formatting requirement: token overhead scaling linearly in the number of categories, validity collapse as the number of categories grows, and ordering bias from the enumeration order. We show that, for tasks where each option can be scored independently, these costs are avoidable through decomposition by option. For tasks that require reasoning over a full distribution (e.g., finding game-theoretic equilibria), second-order prediction is superior. Our work thus demonstrates the complementary qualities of higher-order inference.
Gleb Kuzmin, Lyudmila Rvanova, Ilia Kopanichuk, Vladimir Makharev, Ivan Bakulin, Vladimir Shaposhnikov, Artem Oskin, Petr Anokhin, Aleksandr Nesterov, Artem Vazhentsev, Alexander Panchenko, Elena Tutubalina, Iaroslav Radionovich Bespalov, Dmitry V. Dylov, Artem Shelmanov, Ivan Oseledets. Uncertainty Guardrails for LLM-Based Medical Diagnostic Systems
Large language models (LLMs) are increasingly being adopted in clinical practice, e.g., in medical diagnostic systems. However, their reliability is still a concern, as LLMs are prone to hallucinations and could lead to misdiagnosis, inappropriate treatment decisions, and potential harm to patients. To mitigate these risks, LLM-based medical diagnostic systems should incorporate safety guardrails capable of detecting suspicious predictions and flagging them for human review or further clinical assessment. In this work, we employ uncertainty quantification (UQ) as a safety mechanism within an enterprise diagnostic system. We evaluate 16 UQ methods and propose two task-specific modifications tailored to the medical diagnostic workflow. First, we show that methods based on semantic similarity between generated samples substantially benefit from replacing a general-purpose NLI model with a domain-specific model. Second, we adapt a supervised UQ method that relies on LLM internal states for the detection of erroneous generation from a black-box enterprise LLM accessible only through a text-generation API. Our approach leverages the internal states of an open-source, lightweight proxy white-box model. We demonstrate that this approach outperforms all other UQ methods in the considered setting.
Marian-Sergiu Nistor. ReViB: Mitigating Hallucinations in Large Vision-Language Models via Recovered Visual Bias Contrastive Decoding
Hallucination remains a central failure mode of large vision-language models (LVLMs), which generate descriptions of objects, attributes, or relations absent from the image. A leading explanation is that models fall back on language priors and statistical biases acquired during pretraining rather than grounding their output in the visual input. Contrastive decoding methods counter this by contrasting the model’s prediction against one obtained from a perturbed input, but the perturbation is generic and only loosely aligned with the bias it aims to remove. We propose ReViB (Recovered Visual Bias), a training-free contrastive decoding method that recovers this bias directly and subtracts it. Because an LVLM’s visual encoding captures coarse, prior-aligned semantics more than fine visual detail, we hypothesize that inverting it back into image space through unCLIP reconstructs a scene dominated by the model’s statistical bias. Contrasting the original prediction against this reconstruction removes the recovered bias and yields more grounded outputs. ReViB consistently reduces object and attribute hallucination on the POPE and MME benchmarks, with the largest gains on the popular and adversarial splits where language priors dominate. We further show that, contrary to recent work claiming contrastive decoding applies only a crude, unidirectional shift to the output distribution, ReViB revises predictions in both directions and can lower the yes-rate while improving accuracy.
Ajay Mandyam Rangarajan, Jeyashree Krishnan. LLM Judges Agree With Each Other and Disagree With Human-Grounded Readability Metrics
Text is increasingly written with language models and increasingly assessed by them, which raises a question. When we ask whether a piece of text is well written, what is the reference, the human reader or the model? We study this for readability, where decades of formulas calibrated on human reading performance supply an external, model independent reference against which a judge can be checked. On a longitudinal corpus of 30,595 NeurIPS abstracts from 1987 to 2025 we compute 15 classical readability metrics, and on the 24,772 of them from 1987 to 2024 we run six open-weight instruction-tuned models under three prompt templates, of which 24,754 returned a parseable score. The 15 classical metrics agree that readability declined over the period. All six models under all three prompts report the opposite, drifting without a consistent direction before 2022 and rising in all 18 model and prompt combinations after it. We then measure the reliability signals a practitioner would actually check. The judges are highly repeatable under resampling, between 44\% and 98\% of abstracts receive an identical score on all three sampled generations, and they agree with each other at Spearman 0.37 to 0.41 per paper, giving a panel reliability $\alpha$ of 0.75 to 0.79. That is above the 0.70 conventionally taken as acceptable, so on the usual criteria this panel would be reported as reliable. The direction of its yearly trend is also robust to the standardization baseline. Every one of these checks passes, and all six models still point the opposite way from the reference. Agreement and stability are therefore not evidence of validity. This matters wherever a judge is asked to score a subjective human construct, one defined by its effect on a reader, such as readability, cognitive load, clarity or helpfulness, where no ground truth is available to check the judge against. A per-paper analysis shows the judge score correlates positively with density and sensational language and negatively with sentence length, a partly opposed feature set from the classical reference. If model judgment becomes the operative standard for clear writing, the human reading cost these formulas were built to estimate is no longer measured at any point in the process.
Haoxuan Luo, Jameson Sandler, Nando Fioretto. From Positionwise Confidence to Prefix Scheduling: Verifier Skipping in Speculative Decoding
Speculative decoding is a leading technique to reduce the cost of autoregressive generation by using a small drafter to propose several tokens, which are then verified in parallel by a larger target model. Speculative diffusion decoding (SDD) further removes sequential drafting by generating every position in a draft block in parallel with a discrete diffusion model. However, SDD still invokes the target on every block, leaving verification as a potential bottleneck. This paper recognizes that this creates a new control handle: whether to invoke the verifier at all. Thus, we study verifier skipping, a lossy policy that commits a selected draft prefix directly, and ask which confidence signal should schedule it. Interestingly, our study finds that better token predictors need not yield better schedulers: skips require contiguous high-confidence prefixes, while short skips can induce additional drafting rounds. To study this mismatch, we compare raw confidence with learned marginal and conditional survival scores under the same policy, using Strict SDD, lenience, and top-$k$ acceptance as baselines. On HumanEval with DiffuCoder-7B-Instruct and Qwen3-32B, all three confidence signals save $9.6\%$ to $13.5\%$ of verifier calls at the same observed pass@1 as Strict SDD. Surprisingly, raw confidence saves the most; marginal survival has higher positionwise AUROC than raw confidence at most positions, yet neither learned signal dominates online. Our analysis shows that verifier skipping is a useful new lossy axis and, surprisingly, its key challenge is prefix scheduling rather than token prediction alone.
Harit Vishwakarma, Klara Kaleb, Yee Whye Teh. Don’t Be Choosy: Scoring over Choosing for Verbalized LLM Confidence
Accurate uncertainty estimation is critical for the reliable deployment of large language models (LLMs). A simple and widely used approach is to ask a model to verbalize its confidence on each item, referred to as AbsScore henceforth. However, AbsScores can be poorly discriminative, exhibiting overconfidence or collapsing to a small set of preferred values. Recent work proposed pairwise elicitation as an alternative, asking the model which of two items it is more confident about. While effective, this formulation forces a preference even when confidences are similar and discards information about the magnitude of the confidence difference. To overcome these limitations, we introduce \textbf{PairScore}, a simple alternative: present two items together, but ask the model to score both rather than choose between them. This preserves graded confidence while retaining the benefits of pairwise elicitation, and the resulting scores can be aggregated directly without fitting a separate ranking model. Across diverse models and tasks, including claim verification and best-of-$N$ response selection, PairScore provides a strong alternative to both AbsScore and PairChoice. We further find that PairScore can distinguish items whose AbsScores collapse, while remaining robust to the choice of paired item. Together, these results demonstrate that PairScore offers a simple and effective way to elicit more informative confidence from LLMs.
Sebastian A Nechita, Shubhalaxmi Mukherjee, Pradeep K. Murukannaiah. Conformal Prediction for Complex Fact-Checking with Large Language Models
Large Language Models (LLMs) are increasingly deployed in high-stakes applications, where unreliable outputs can have serious consequences. However, quantifying their uncertainty and providing verifiable guarantees on their predictions remains an open challenge. In this paper, we investigate the use of conformal prediction (CP) to address this gap in LLM-based fact-checking. We evaluate three non-conformity scores on the claim verification task by formulating it as a multiple-choice question answering (MCQA) problem. We assess the LLMs’ confidence in classifying claims into a six-class rating system. We find that access to the relevant evidence sharply reduced prediction set sizes, yielding sets small enough to be useful to a human fact-checker. Without access to external evidence, models could not reliably distinguish true from false claims on their own at the 90\% marginal coverage rate. This paper advances the application of LLMs in knowledge-intensive tasks where known error bounds in the model’s response is essential for reliability.
Davide Mazzaccara, Leonardo Bertolazzi, Raffaella Bernardi. Different Facets of Verbalised Overconfidence: an Interpretability Study
Large language models tend to be overconfident, giving assertive answers when the evidence suggests hedging or abstention. Using controlled reasoning scenarios that manipulate logical necessity and possibility, we study this behavior in Qwen3-4B and Qwen3-8B. At the interpretability level, we propose a method that differentially identifies transcoder features responsible for uncertainty and certainty outputs. Our analysis reveals Qwen3’s default mechanism favors certainty generation through a broad coalition of shared features, while uncertainty is implemented as a sparse override mediated by a small set of dedicated features. Intervening on these uncertainty features both causally proves this imbalance underlying overconfidence and also mitigate overconfident errors. The same set of uncertainty features generalise across other ways of expressing uncertainty (numeric scores, abstention rate), languages and to a different task.
Marina Ananyan, Sultan Akimaliyev, Peter Blohm, Vikas K Garg. Confidence-Weighted Conformal Calibration with Untrusted Factuality Scores
Large language models can produce incorrect responses with high confidence, making it important to quantify uncertainty in their outputs. Conformal prediction constructs prediction sets with a specified coverage guarantee, and split conformal obtains this guarantee from a separate calibration set without changing the underlying model. However, obtaining calibration scores can be costly when they require human annotations. Automatic judges can be a cheaper source of additional scores, but naively pooling them with human scores can invalidate the coverage guarantee. We propose confidence weighted pooling, which uses automatic scores to adjust the calibration threshold while preserving a chosen minimum coverage level. The method compares the trusted and automatic score distributions while accounting for uncertainty from limited data. Stronger evidence of similarity allows the automatic scores to have more influence on the threshold. Trusted scores determine how far the threshold can be lowered while preserving marginal coverage of at least (c_{\min}), assuming exchangeability between the trusted calibration and test data. In controlled experiments with a standard coverage target of (90\%) and a protected minimum of (c_{\min}=85\%), the method achieved a mean normalized threshold reduction of (0.241), measured relative to the interval between the standard threshold and the trusted floor. In the experiments on RAGTruth and DialFact, the pooled thresholds were not lower than the standard thresholds, so the method returned ordinary split conformal.
Alessandra Terranova, Alexandra Birch, Björn Ross. Internal, Verbalised, and Perceived Confidence in Multilingual LLMs
As large language models (LLMs) are increasingly used by humans across multiple languages, there is a need for models to express their uncertainty in a way that is well-calibrated, honest, and aligned to how users interpret it. We construct the first multilingual and distributional dataset of LLM linguistic expressions of confidence, spanning seven languages. Each expression is labelled with a corresponding probability distribution and the belief intervals labelled by native-speaker annotators. We argue that human understanding of linguistic hedges and expressions of confidence varies cross-linguistically, and it is fuzzy and at times vague, which makes the use of point estimates for confidence insufficient. We study expressions of confidence in LLMs both as linguistic and numerical verbalisations, and as internal token-probability and and consistency-based signals. We find that, in the multiple-choice-question setting, across elicitation methods, English is better calibrated than the other languages studied, and that models lose calibration when attempting to surface uncertainty, with logit extraction achieving the best calibration, followed by consistency-based methods and linguistic confidence, while verbalised numerical confidence exhibits the poorest calibration. Additionally, instruction tuning enhances discrimination but degrades calibration for internal methods, whereas it consistently improves calibration for verbalisation-based methods.
Maiya Goloburda, Roman Vashurin, Fedor Chernogorskii, Nurkhan Laiyk, Daniil Orel, Preslav Nakov, Maxim Panov. Why Don’t You Know? Evaluating the Impact of Uncertainty Sources on Uncertainty Quantification in LLMs
As Large Language Models (LLMs) are increasingly deployed in real-world applications, reliable uncertainty quantification (UQ) becomes critical for safe and effective use. Most existing UQ approaches for language models aim to produce a single confidence score – for example, estimating the probability that a model’s answer is correct. However, uncertainty in natural language tasks arises from multiple distinct sources, including model knowledge gaps, output variability, and input ambiguity, which have different implications for system behavior and user interaction. In this work, we study how the source of uncertainty impacts the behavior and effectiveness of existing UQ methods. To enable controlled analysis, we construct a human-validated dataset that introduces distinct uncertainty sources in a controlled way, enabling the first systematic evaluation of how uncertainty source affects UQ method behavior. Our experiments reveal that method families fail in predictable, source-specific ways, and that no single method reliably handles all sources. These findings highlight the need for uncertainty-aware methods that explicitly account for the source of uncertainty in large language models. Our data and code are available at \url{http://anonymous.for.review}.