LLM Uncertainty Estimation

Uncertainty quantification (UQ) methods for large language models are predominantly designed by hand based on domain knowledge and heuristics, limiting their scalability and generality. We apply LLM-powered evolutionary search to automatically discover unsupervised UQ methods represented as Python programs. On the task of atomic claim verification, our evolved methods outperform strong manually-designed baselines, achieving up to 6.7% relative ROC-AUC improvement across 9 datasets while generalizing robustly out-of-distribution. Qualitative analysis reveals that different LLMs employ qualitatively distinct evolutionary strategies: Claude models consistently design high-feature-count linear estimators, while Gpt-oss-120B gravitates toward simpler and more interpretable positional weighting schemes. Surprisingly, only Sonnet 4.5 and Opus 4.5 reliably leverage increased method complexity to improve performance -- Opus 4.6 shows an unexpected regression relative to its predecessor. Overall, our results indicate that LLM-powered evolutionary search is a promising paradigm for automated, interpretable hallucination detector design.

When deploying large language models (LLMs) to safety-critical applications, uncertainty quantification (UQ) is of utmost importance to self-assess the reliability of the LLM-based decisions. However, such decisions typically suffer from overconfidence, particularly after parameter-efficient fine-tuning (PEFT) for downstream domain-specific tasks with limited data. Existing methods to alleviate this issue either rely on Laplace approximation based post-hoc framework, which may yield suboptimal calibration depending on the training trajectory, or variational Bayesian training that requires multiple complete forward passes through the entire LLM backbone at inference time for Monte Carlo estimation, posing scalability challenges for deployment. To address these limitations, we build on the Bayesian last layer (BLL) model, where the LLM-based deterministic feature extractor is followed by random last layer parameters for uncertainty reasoning. Since existing low-rank adapters (LoRA) for PEFT have limited expressiveness due to rank collapse, we address this with Polar-decomposed Low-rank Adapter Representation (PoLAR), an orthogonalized parameterization paired with Riemannian optimization to enable more stable and expressive adaptation. Building on this PoLAR-BLL model, we leverage the variational (V) inference framework to put forth a scalable Bayesian fine-tuning approach which jointly seeks the PoLAR parameters and approximate posterior of the last layer parameters via alternating optimization. The resulting PoLAR-VBLL is a flexible framework that nicely integrates architecture-enhanced optimization with scalable Bayesian inference to endow LLMs with well-calibrated UQ. Our empirical results verify the effectiveness of PoLAR-VBLL in terms of generalization and uncertainty estimation on both in-distribution and out-of-distribution data for various common-sense reasoning tasks.

Detecting whether a model's internal knowledge is sufficient to correctly answer a given question is a fundamental challenge in deploying responsible LLMs. In addition to verbalising the confidence by LLM self-report, more recent methods explore the model internals, such as the hidden states of the response tokens, to capture how much knowledge is activated. We argue that such activated knowledge may not align with what the query requires, e.g., capturing the stylistic and length-related features that are uninformative for answering the query. To fill the gap, we propose GRADE (GRAdient Dynamics for knowlEdge gap detection), which quantifies the knowledge gap via the cross-layer rank ratio of the gradient to that of the corresponding hidden state subspace. This is motivated by the property of gradients as estimators of the required knowledge updates for a given target. We validate GRADE on six benchmarks, demonstrating its effectiveness and robustness to input perturbations. In addition, we present a case study demonstrating how the gradient chain can generate interpretable explanations of knowledge gaps in long-form answers. Code is available at https://github.com/yjEugenia/llm-aware

Large language models (LLMs) have been proposed as alternatives to human experts for estimating unknown quantities with associated uncertainty, a process known as Bayesian elicitation. We test this by asking eleven LLMs to estimate population statistics, such as health prevalence rates, personality trait distributions, and labor market figures, and to express their uncertainty as 95% credible intervals. We vary each model's reasoning effort (low, medium, high) to test whether more "thinking" improves results. Our findings reveal three key results. First, larger, more capable models produce more accurate estimates, but increasing reasoning effort provides no consistent benefit. Second, all models are severely overconfident: their 95% intervals contain the true value only 9--44% of the time, far below the expected 95%. Third, a statistical recalibration technique called conformal prediction can correct this overconfidence, expanding the intervals to achieve the intended coverage. In a preliminary experiment, giving models web search access degraded predictions for already-accurate models, while modestly improving predictions for weaker ones. Models performed well on commonly discussed topics but struggled with specialized health data. These results indicate that LLM uncertainty estimates require statistical correction before they can be used in decision-making.

While Large Language Models (LLMs) have demonstrated exceptional proficiency in code completion, they typically adhere to a Hard Completion (HC) paradigm, compelling the generation of fully concrete code even amidst insufficient context. Our analysis of 3 million real-world interactions exposes the limitations of this strategy: 61% of the generated suggestions were either edited after acceptance or rejected despite exhibiting over 80% similarity to the user's subsequent code, suggesting that models frequently make erroneous predictions at specific token positions. Motivated by this observation, we propose Adaptive Placeholder Completion (APC), a collaborative framework that extends HC by strategically outputting explicit placeholders at high-entropy positions, allowing users to fill directly via IDE navigation. Theoretically, we formulate code completion as a cost-minimization problem under uncertainty. Premised on the observation that filling placeholders incurs lower cost than correcting errors, we prove the existence of a critical entropy threshold above which APC achieves strictly lower expected cost than HC. We instantiate this framework by constructing training data from filtered real-world edit logs and design a cost-based reward function for reinforcement learning. Extensive evaluations across 1.5B--14B parameter models demonstrate that APC reduces expected editing costs from 19% to 50% while preserving standard HC performance. Our work provides both a theoretical foundation and a practical training framework for uncertainty-aware code completion, demonstrating that adaptive abstention can be learned end-to-end without sacrificing conventional completion quality.

We identify a previously unknown calibration property of large language models: providing one confirmed intermediate fact toward a multi-step reasoning chain increases the model's confident-wrong-answer rate before full evidence eliminates it. We call this anchored confabulation: a partial anchor commits the model to confident parametric completion of remaining reasoning steps. We formalize it as Parametric Hallucination Confidence (PHC) and establish it across six lines of evidence including a causal injection experiment (PHC 0.613 to 0.656 to 0.595 to 0.536, N=160) and capability scaling across five model families (Spearman rho=0.900, p=0.037). The Anchoring Threshold Law k*(n)=floor(n/3) predicts PHC amplification by hop depth with four confirmed predictions. Applied to RAG routing, a LearnedRouter exploiting PHC closes 81.1% of the oracle performance gap (macro F1=0.426, p<1e-6) on 1,800 queries across four benchmarks with no model fine-tuning and 50x fewer labels than prior RL-based work. An epistemic humility prompt reduces the PHC spike by -0.118; explicit self-rating (PHC=0.684, p<0.001) outperforms lexical confidence as a routing signal.

To schedule LLM inference, the shortest job first (SJF) principle is favorable by prioritizing requests with short output lengths to avoid head-of-line (HOL) blocking. Existing methods usually predict a single output length for each request to facilitate scheduling. We argue that such a point estimate does not match the stochastic decoding process of LLM inference, where output length is uncertain by nature and determined by when the end-of-sequence (EOS) token is sampled. Hence, the output length of each request should be fitted with a distribution rather than a single value. With an in-depth analysis of empirical data and the stochastic decoding process, we observe that output length follows a heavy-tailed distribution and can be fitted with the log-t distribution. On this basis, we propose a simple metric called Tail Inflated Expectation (TIE) to replace the output length in SJF scheduling, which adjusts the expectation of a log-t distribution with its tail probabilities to account for the risk that a request generates long outputs. To evaluate our TIE scheduler, we compare it with three strong baselines, and the results show that TIE reduces the per-token latency by $2.31\times$ for online inference and improves throughput by $1.42\times$ for offline data generation.

Uncertainty estimation (UE) aims to quantify uncertainty in large language model (LLM) outputs to flag potential hallucinations and support reliable deployment. However, UE metrics often exhibit unstable performance across models and datasets, limiting their practical utility. We formalise this phenomenon as proxy failure, arising because UE metrics are derived from model behaviour rather than being explicitly grounded in factual correctness. We show that when mutual information between a UE score and correctness is low, the score's discriminability must be close to chance. To address this, we propose Truth AnChoring (TAC), a lightweight post-hoc calibration method that maps raw UE scores to truth-aligned correctness probabilities. Even with noisy and few-shot supervision, TAC continues to improve calibration, providing a practical calibration protocol. Our findings highlight the limitations of treating heuristic UE metrics as direct indicators of uncertainty about factual correctness, and position our TAC as a principled step toward more reliable uncertainty estimation for LLMs. The code repository is available at https://github.com/ponhvoan/TruthAnchor/.

A robust decision-making process must take into account uncertainty, especially when the choice involves inherent risks. Because artificial intelligence (AI) systems are increasingly integrated into decision-making workflows, managing uncertainty relies more and more on the metacognitive capabilities of these systems; i.e, their ability to assess the reliability of and regulate their own decisions. Hence, it is crucial to employ robust methods to measure the metacognitive abilities of AI. This paper is primarily a methodological contribution that highlights a key limitation of commonly used measures of AI metacognitive sensitivity--the ability to generate confidence ratings that distinguish correct from incorrect responses. We then draw attention to the meta-d' framework, a well-established approach from psychology and neuroscience designed to address this limitation. Moreover, we propose to leverage signal detection theory (SDT) to measure the ability of AIs to spontaneously regulate their decisions based on uncertainty and risk. To demonstrate the practical utility of these psychophysical frameworks, we conduct two series of experiments on three large language models (LLMs)--GPT-5, DeepSeek-V3.2-Exp, and Mistral-Medium-2508.

Existing methods for quantifying predictive uncertainty in neural networks are either computationally intractable for large language models or require access to training data that is typically unavailable. We derive a lightweight alternative through two approximations: a first-order Taylor expansion that expresses uncertainty in terms of the gradient of the prediction and the parameter covariance, and an isotropy assumption on the parameter covariance. Together, these yield epistemic uncertainty as the squared gradient norm and aleatoric uncertainty as the Bernoulli variance of the point prediction, from a single forward-backward pass through an unmodified pretrained model. We justify the isotropy assumption by showing that covariance estimates built from non-training data introduce structured distortions that isotropic covariance avoids, and that theoretical results on the spectral properties of large networks support the approximation at scale. Validation against reference Markov Chain Monte Carlo estimates on synthetic problems shows strong correspondence that improves with model size. We then use the estimates to investigate when each uncertainty type carries useful signal for predicting answer correctness in question answering with large language models, revealing a benchmark-dependent divergence: the combined estimate achieves the highest mean AUROC on TruthfulQA, where questions involve genuine conflict between plausible answers, but falls to near chance on TriviaQA's factual recall, suggesting that parameter-level uncertainty captures a fundamentally different signal than self-assessment methods.

We train a 318M-parameter Transformer language model from scratch on a curated corpus of 1.56 billion tokens of pure Classical Chinese, with zero English characters or Arabic numerals. Through systematic out-of-distribution (OOD) testing, we ask whether the model distinguishes known from unknown inputs, and whether it expresses that distinction in its generated text. We find a clear dissociation between internal and external uncertainty. Internally, the model exhibits a perplexity jump ratio of 2.39x between real and fabricated historical events (p = 8.9e-11, n = 92 per group), with semi-fabricated events (real figures + fictional actions) showing the highest perplexity (4.24x, p = 1.1e-16), demonstrating genuine factual encoding beyond syntactic pattern matching. Externally, however, the model never learns to express uncertainty: classical Chinese epistemic markers appear at lower rates for OOD questions (3.5%) than in-distribution ones (8.3%, p = 0.023), reflecting rhetorical conventions in the training data rather than genuine metacognition. We test both findings across three languages (Classical Chinese, English, Japanese), three writing systems, and eight models from 110M to 1.56B. The internal factual-encoding effect replicates in six of the eight models, emerging with scale (the two smallest Japanese models do not yet separate real from fabricated history), while the external absence of uncertainty expression holds across all eight. We further show that uncertainty expression frequency is determined entirely by training data conventions -- not epistemic states -- with Classical Chinese models showing a "humility paradox" (more hedging for known topics), while Japanese models almost never hedge. We argue that metacognitive expression -- the ability to say "I don't know" -- does not emerge from language modeling alone and requires explicit training signals such as RLHF.

Routing is widely used to scale large language models, from Mixture-of-Experts gating to multi-model/tool selection. A common belief is that routing to a task ``expert'' activates sparser internal computation and thus yields more certain and stable outputs (the Sparsity--Certainty Hypothesis). We test this belief by injecting routing-style meta prompts as a textual proxy for routing signals in front of frozen instruction-tuned LLMs. We quantify (C1) internal density via activation sparsity, (C2) domain-keyword attention, and (C3) output stability via predictive entropy and semantic variation. On a RouterEval subset with three instruction-tuned models (Qwen3-8B, Llama-3.1-8B-Instruct, and Mistral-7B-Instruct-v0.2), meta prompts consistently densify early/middle-layer representations rather than increasing sparsity; natural-language expert instructions are often stronger than structured tags. Attention responses are heterogeneous: Qwen/Llama reduce keyword attention, while Mistral reinforces it. Finally, the densification--stability link is weak and appears only in Qwen, with near-zero correlations in Llama and Mistral. We present RIDE as a diagnostic probe for calibrating routing design and uncertainty estimation.

Uncertainty estimation in multi-LLM systems remains largely single-model-centric: existing methods quantify uncertainty within each model but do not adequately capture semantic disagreement across models. To address this gap, we propose Collaborative Entropy (CoE), a unified information-theoretic metric for semantic uncertainty in multi-LLM collaboration. CoE is defined on a shared semantic cluster space and combines two components: intra-model semantic entropy and inter-model divergence to the ensemble mean. CoE is not a weighted ensemble predictor; it is a system-level uncertainty measure that characterizes collaborative confidence and disagreement. We analyze several core properties of CoE, including non-negativity, zero-value certainty under perfect semantic consensus, and the behavior of CoE when individual models collapse to delta distributions. These results clarify when reducing per-model uncertainty is sufficient and when residual inter-model disagreement remains. We also present a simple CoE-guided, training-free post-hoc coordination heuristic as a practical application of the metric. Experiments on TriviaQA and SQuAD with LLaMA-3.1-8B-Instruct, Qwen-2.5-7B-Instruct, and Mistral-7B-Instruct show that CoE provides stronger uncertainty estimation than standard entropy- and divergence-based baselines, with gains becoming larger as additional heterogeneous models are introduced. Overall, CoE offers a useful uncertainty-aware perspective on multi-LLM collaboration.

Transformers for language modeling usually rely on deterministic internal computation, with uncertainty expressed mainly at the output layer. We introduce variational neurons into Transformer feed-forward computation so that uncertainty becomes part of the internal computation itself. Concretely, we replace deterministic feed-forward units with local variational units based on EVE while preserving the overall Transformer backbone. We evaluate this design in compact next-token language-modeling settings. We compare deterministic and variational variants with both predictive and probabilistic criteria. Alongside negative log-likelihood, perplexity and accuracy, we analyze calibration, conditional variance, mutual information and latent-usage statistics. The resulting picture is clear. Variational neurons integrate stably into Transformers, preserve strong predictive performance and produce informative uncertainty signals. The experiments also show that task quality, useful depth and internal stability are distinct properties. These results establish variational Transformers as a practical form of uncertainty-aware language modeling. They show that Transformers can predict with an explicit internal structure of uncertainty, which supports stronger probabilistic evaluation and a more informative analysis of model behavior.

The generation of texts using Large Language Models (LLMs) is inherently uncertain, with sources of uncertainty being not only the generation of texts, but also the prompt used and the downstream interpretation. Within this work, we provide a formal framework for the measurement of uncertainty that takes these different aspects into account. Our framework models prompting, generation, and interpretation as interconnected autoregressive processes that can be combined into a single sampling tree. We introduce filters and objective functions to describe how different aspects of uncertainty can be expressed over the sampling tree and demonstrate how to express existing approaches towards uncertainty through these functions. With our framework we show not only how different methods are formally related and can be reduced to a common core, but also point out additional aspects of uncertainty that have not yet been studied.

Accurate uncertainty quantification is crucial for making reliable decisions in various supervised learning scenarios, particularly when dealing with complex, multimodal data such as images and text. Current approaches often face notable limitations, including rigid assumptions and limited generalizability, constraining their effectiveness across diverse supervised learning tasks. To overcome these limitations, we introduce Generative Score Inference (GSI), a flexible inference framework capable of constructing statistically valid and informative prediction and confidence sets across a wide range of multimodal learning problems. GSI utilizes synthetic samples generated by deep generative models to approximate conditional score distributions, facilitating precise uncertainty quantification without imposing restrictive assumptions about the data or tasks. We empirically validate GSI's capabilities through two representative scenarios: hallucination detection in large language models and uncertainty estimation in image captioning. Our method achieves state-of-the-art performance in hallucination detection and robust predictive uncertainty in image captioning, and its performance is positively influenced by the quality of the underlying generative model. These findings underscore the potential of GSI as a versatile inference framework, significantly enhancing uncertainty quantification and trustworthiness in multimodal learning.

As Large Language Model (LLM) agents are increasingly deployed in open-ended domains like software engineering, they frequently encounter underspecified instructions that lack crucial context. While human developers naturally resolve underspecification by asking clarifying questions, current agents are largely optimized for autonomous execution. In this work, we systematically evaluate the clarification-seeking abilities of LLM agents on an underspecified variant of SWE-bench Verified. We propose an uncertainty-aware multi-agent scaffold that decouples underspecification detection from code execution. Across both proprietary and open-weight frontier LLMs, our scaffold achieves a 69.40% task resolve rate, significantly outperforming a standard single-agent setup and closing the performance gap with agents operating on fully specified instructions. Furthermore, we find that the multi-agent system exhibits well-calibrated information-seeking behavior, conserving queries on simple tasks while proactively seeking information on more complex issues. These findings indicate that current models can be turned into proactive collaborators, where agents independently recognize when to ask questions to elicit missing information in real-world, underspecified tasks.

Standard evaluation of LLM confidence relies on calibration metrics (ECE, Brier score) that conflate two capacities: how much a model knows (Type-1 accuracy) and how well its confidence signal tracks that knowledge (Type-2 metacognitive sensitivity). We apply Signal Detection Theory to decompose them, treating token-level normalised log-probability as a graded confidence variable and answer correctness as the state to be discriminated. We characterise the Type-2 ROC of this signal, including its unequal-variance structure via z-ROC analysis, and -- because the meta-d' efficiency ratio is not well defined for open-ended QA, which lacks a two-alternative Type-1 decision -- quantify efficiency with a model-free information measure, normalised metacognitive information (meta-I_2r). Across four LLMs and 224,000 factual QA trials we find: (1) metacognitive information varies by a factor of 1.98 across models and is not predicted by accuracy, the rank correlation being -0.80 on TriviaQA and +0.00 on Natural Questions; (2) the confidence signal has model-specific unequal-variance structure (z-ROC slopes 0.78 to 1.18) invisible to calibration metrics, the slope ordering replicating on NQ; (3) efficiency is domain-specific, weakest in Science & Technology for every model; (4) temperature dissociates accuracy from metacognitive information, which stays near-flat for three of four models while accuracy falls; and (5) metacognitive information tracks the accuracy gain from confidence-based abstention exactly (rho = +1.00) while accuracy does not. All estimates carry permutation nulls and bootstrap confidence intervals. This version (v3) corrects a differential length bias in the automated correctness scorer, validated against 1,830 human adjudications; the inverse accuracy-efficiency coupling reported in v1 and v2 does not survive relabelling. See the version note on page 1.

Understanding why a large language model (LLM) is uncertain about the response is important for their reliable deployment. Current approaches, which either provide a single uncertainty score or rely on the classical aleatoric-epistemic dichotomy, fail to offer actionable insights for improving the generative model. Recent studies have also shown that such methods are not enough for understanding uncertainty in LLMs. In this work, we advocate for an uncertainty decomposition framework that dissects LLM uncertainty into three distinct semantic components: (i) input ambiguity, arising from ambiguous prompts; (ii) knowledge gaps, caused by insufficient parametric evidence; and (iii) decoding randomness, stemming from stochastic sampling. Through a series of experiments we demonstrate that the dominance of these components can shift across model size and task. Our framework provides a better understanding to audit LLM reliability and detect hallucinations, paving the way for targeted interventions and more trustworthy systems.

Understanding and quantifying uncertainty in large language model (LLM) outputs is critical for reliable deployment. However, traditional evaluation approaches provide limited insight into model confidence at individual token positions during generation. To address this issue, we introduce LogitScope, a lightweight framework for analyzing LLM uncertainty through token-level information metrics computed from probability distributions. By measuring metrics such as entropy and varentropy at each generation step, LogitScope reveals patterns in model confidence, identifies potential hallucinations, and exposes decision points where models exhibit high uncertainty, all without requiring labeled data or semantic interpretation. We demonstrate LogitScope's utility across diverse applications including uncertainty quantification, model behavior analysis, and production monitoring. The framework is model-agnostic, computationally efficient through lazy evaluation, and compatible with any HuggingFace model, enabling both researchers and practitioners to inspect LLM behavior during inference.

RLHF-aligned language models exhibit response homogenization: on TruthfulQA (n=790), 40-79% of questions produce a single semantic cluster across 10 i.i.d. samples. On affected questions, sampling-based uncertainty methods have zero discriminative power (AUROC=0.500), while free token entropy retains signal (0.603). This alignment tax is task-dependent: on GSM8K (n=500), token entropy achieves 0.724 (Cohen's d=0.81). A base-vs-instruct ablation confirms the causal role of alignment: the base model shows 1.0% single-cluster rate vs. 28.5% for the instruct model (p < 10^{-6}). A training stage ablation (Base 0.0% -> SFT 1.5% -> DPO 4.0% SCR) localizes the cause to DPO, not SFT. Cross-family replication on four model families reveals alignment tax severity varies by family and scale. We validate across 22 experiments, 5 benchmarks, 4 model families, and 3 model scales (3B-14B), with Jaccard, embedding, and NLI-based baselines at three DeBERTa scales (all ~0.51 AUROC). Cross-embedder validation with two independent embedding families rules out coupling bias. Cross-dataset validation on WebQuestions (58.0% SCR) confirms generalization beyond TruthfulQA. The central finding -- response homogenization -- is implementation-independent and label-free. Motivated by this diagnosis, we explore a cheapest-first cascade (UCBD) over orthogonal uncertainty signals. Selective prediction raises GSM8K accuracy from 84.4% to 93.2% at 50% coverage; weakly dependent boundaries (|r| <= 0.12) enable 57% cost savings.

Large language models (LLMs) inherently operate over a large generation space, yet conventional usage typically reports the most likely generation (MLG) as a point prediction, which underestimates the model's capability: although the top-ranked response can be incorrect, valid answers may still exist within the broader output space and can potentially be discovered through repeated sampling. This observation motivates moving from point prediction to set-valued prediction, where the model produces a set of candidate responses rather than a single MLG. In this paper, we propose a principled framework for set-valued prediction, which provides feasibility-aware coverage guarantees. We show that, given the finite-sampling nature of LLM generation, coverage is not always achievable: even with multiple samplings, LLMs may fail to yield an acceptable response for certain questions within the sampled candidate set. To address this, we establish a minimum achievable risk level (MRL), below which statistical coverage guarantees cannot be satisfied. Building on this insight, we then develop a data-driven calibration procedure that constructs prediction sets from sampled responses by estimating a rigorous threshold, ensuring that the resulting set contains a correct answer with a desired probability whenever the target risk level is feasible. Extensive experiments on six language generation tasks with five LLMs demonstrate both the statistical validity and the predictive efficiency of our framework.

While retrieval-augmented generation (RAG) enhances LLM performance, it does not eliminate hallucinations, making accurate detection essential. Uncertainty-based methods are attractive for this purpose because they can be integrated into real-world pipelines with little overhead. One of the most widely used uncertainty signals is predictive entropy. We show, however, that entropy can be unreliable in RAG settings and trace this limitation to two opposing internal effects. Induction heads, which copy patterns from earlier context, causally support correct responses and lower predictive entropy, but they also appear to co-activate entropy neurons that push it back up. As a result, correct, context-grounded responses can still receive high uncertainty scores. To address this, we propose INTRYGUE (Induction-Aware Entropy Gating for Uncertainty Estimation), a training-free, mechanistically grounded method that gates predictive entropy by an attention-based estimate of induction-head activity. Evaluated across four RAG-style benchmarks and six open-source LLMs (4B to 13B parameters), INTRYGUE performs competitively against a wide range of baselines, matching or exceeding the strongest of them in most settings. Our findings suggest that hallucination detection in RAG benefits from combining predictive uncertainty with interpretable internal signals of context utilization.

Selective prediction systems can mitigate harms resulting from language model hallucinations by abstaining from answering in high-risk cases. Uncertainty quantification techniques are often employed to identify such cases, but are rarely evaluated in the context of the wider selective prediction policy and its ability to operate at low target error rates. We identify a model-dependent failure mode of entropy-based uncertainty methods that leads to unreliable abstention behaviour, and address it by combining entropy scores with a correctness probe signal. We find that across three QA benchmarks (TriviaQA, BioASQ, MedicalQA) and four model families, the combined score generally improves both the risk--coverage trade-off and calibration performance relative to entropy-only baselines. Our results highlight the importance of deployment-facing evaluation of uncertainty methods, using metrics that directly reflect whether a system can be trusted to operate at a stated risk level.

Multi-agent LLM systems, where multiple prompted instances of a language model independently answer questions, are increasingly used for complex reasoning tasks. However, existing methods for quantifying the uncertainty of their collective outputs rely on shallow voting statistics that discard the rich semantic information in agents' reasoning. We introduce DiscoUQ, a framework that extracts and leverages the structure of inter-agent disagreement -- both linguistic properties (evidence overlap, argument strength, divergence depth) and embedding geometry (cluster distances, dispersion, cohesion) -- to produce well-calibrated confidence estimates. We propose three methods of increasing complexity: DiscoUQ-LLM (logistic regression on LLM-extracted structure features), DiscoUQ-Embed (logistic regression on embedding geometry), and DiscoUQ-Learn (a neural network combining all features). Evaluated on four diverse benchmarks (StrategyQA, MMLU, TruthfulQA, ARC-Challenge) with a 5-agent system using Qwen3.5-27B, DiscoUQ-LLM achieves an average AUROC of 0.802, outperforming the best baseline (LLM Aggregator, 0.791) while being substantially better calibrated (ECE 0.036 vs. 0.098). The learned features generalize across benchmarks with near-zero performance degradation and provide the largest improvements where they are most needed: in the ambiguous "weak disagreement" tier where simple vote counting fails.

Large language models (LLMs) have demonstrated remarkable capabilities across diverse tasks. However, the truthfulness of their outputs is not guaranteed, and their tendency toward overconfidence further limits reliability. Uncertainty quantification offers a promising way to identify potentially unreliable outputs, but most existing methods rely on repeated sampling or auxiliary models, introducing substantial computational overhead. To address these limitations, we propose Semantic Token Clustering (STC), an efficient uncertainty quantification method that leverages the semantic information inherently encoded in LLMs. Specifically, we group tokens into semantically consistent clusters using embedding clustering and prefix matching, and quantify uncertainty based on the probability mass aggregated over the corresponding semantic cluster. Our approach requires only a single generation and does not depend on auxiliary models. Experimental results show that STC achieves performance comparable to state-of-the-art baselines while substantially reducing computational overhead.

Uncertainty estimation is critical for deploying reasoning language models, yet remains poorly understood under extended chain-of-thought reasoning. We study parallel sampling as a fully black-box approach using verbalized confidence and self-consistency. Across three reasoning models and 17 tasks spanning mathematics, STEM, and humanities, we characterize how these signals scale. Both self-consistency and verbalized confidence scale in reasoning models, but self-consistency exhibits lower initial discrimination and lags behind verbalized confidence under moderate sampling. Most uncertainty gains, however, arise from signal combination: with just two samples, a hybrid estimator improves AUROC by up to $+12$ on average and already outperforms either signal alone even when scaled to much larger budgets, after which returns diminish. These effects are domain-dependent: in mathematics, the native domain of RLVR-style post-training, reasoning models achieve higher uncertainty quality and exhibit both stronger complementarity and faster scaling than in STEM or humanities.

Understanding uncertainty in chain-of-thought reasoning is critical for reliable deployment of large language models. In this work, we propose a simple yet effective diagnostic approach based on trajectory shape rather than scalar magnitude. We show that this signal is practical, interpretable, and inexpensive to obtain in black-box settings, while remaining robust across models and datasets. Through extensive ablations and cross-domain replications, we demonstrate its utility for selective prediction and triage. Our findings offer a generalizable insight into uncertainty dynamics in reasoning tasks, with particular focus on numeric and discrete-answer settings.

Long-context inference remains challenging for large language models due to attention dilution and out-of-distribution degradation. Context selection mitigates this limitation by attending to a subset of key-value cache entries, yet most methods allocate a fixed context budget throughout decoding despite highly non-uniform token-level contextual demands. To address this issue, we propose Uncertainty-Triggered Adaptive Context Allocation (UT-ACA), an inference-time framework that dynamically adjusts the context window based on token-wise uncertainty. UT-ACA learns an uncertainty detector that combines semantic embeddings with logit-based confidence while accounting for uncertainty accumulation across decoding steps. When insufficient evidence is indicated, UT-ACA selectively rolls back, expands the context window, and regenerates the token with additional support. Experiments show that UT-ACA substantially reduces average context usage while preserving generation quality in long-context settings. Code and dataset are available at https://github.com/Tommy307/UT-ACA.

Transformer-based language models are widely deployed for reasoning, yet their behavior under inference-time stochasticity remains underexplored. While dropout is common during training, its inference-time effects via Monte Carlo sampling lack systematic evaluation across architectures, limiting understanding of model reliability in uncertainty-aware applications. This work analyzes dropout-induced variability across 19 transformer models using MC Dropout with 100 stochastic forward passes per sample. Dropout robustness is defined as maintaining high accuracy and stable predictions under stochastic inference, measured by standard deviation of per-run accuracies. A cognitive decomposition framework disentangles performance into memory and reasoning components. Experiments span five dropout configurations yielding 95 unique evaluations on 1,000 samples. Results reveal substantial architectural variation. Smaller models demonstrate perfect prediction stability while medium-sized models exhibit notable volatility. Mid-sized models achieve the best overall performance; larger models excel at memory tasks. Critically, 53% of models suffer severe accuracy degradation under baseline MC Dropout, with task-specialized models losing up to 24 percentage points, indicating unsuitability for uncertainty quantification in these architectures. Asymmetric effects emerge: high dropout reduces memory accuracy by 27 percentage points while reasoning degrades only 1 point, suggesting memory tasks rely on stable representations that dropout disrupts. 84% of models demonstrate memory-biased performance. This provides the first comprehensive MC Dropout benchmark for transformers, revealing dropout robustness is architecture-dependent and uncorrelated with scale. The cognitive profiling framework offers actionable guidance for model selection in uncertainty-aware applications.